[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_266":267,"_268":269},"routes/_app.x.$handle.$name",{"_100":99,"_32":44,"_109":90,"_27":54,"_13":14,"_110":-5,"_11":111,"_112":113,"_114":115,"_116":117,"_118":119,"_20":-5,"_120":-5,"_121":-5,"_122":25,"_123":124,"_125":59,"_126":-5,"_127":-5,"_128":129},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_42":43,"_32":44,"_37":45,"_46":-5,"_47":25,"_48":49,"_50":49,"_51":52,"_53":54,"_18":55,"_56":57,"_58":59,"_60":-5,"_61":-5,"_62":63,"_64":65,"_66":67,"_68":69,"_70":71,"_72":-5,"_73":74,"_75":74,"_76":77,"_78":77,"_79":77,"_80":74,"_81":74,"_82":83,"_84":85,"_86":87,"_88":25},"version",79,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":23,"_24":25,"_26":25,"_27":28,"_29":28},"file",{"_30":31,"_32":28,"_33":20,"_34":36,"_37":38,"_39":14,"_40":41,"_24":25},"previous",{"_30":31,"_32":28,"_33":20,"_34":35},"deleted",false,"deletedHere","path","README.md","latestPath","fileId","019ec890-2328-734f-a4be-e3c03c1985c4","name","fileType","content","# call-my-human — send an agent to phone your human 📞\n\nA **public remote MCP server**: connect it to any AI agent and it can dispatch an *autonomous* voice agent to phone your human, **run the whole conversation itself**, hang up, and hand back a structured report. Brief it once (`objective` + `context` + `questions`); it talks, ends the call, and you collect the result. The agent has a real LLM brain running **inside Vapi's stack**, so the call is low-latency and natural — you don't drive it turn by turn. **BYOK: no env vars, no server-side keys, no signup.**\n\n> Siblings, same codebase, three control models: [`call-my-human-no-middle-agent`](https://www.val.town/x/dcm31/call-my-human-no-middle-agent) is the *puppet* (you drive every word — full control, high latency); [`dispatch-my-human-escape`](https://www.val.town/x/dcm31/dispatch-my-human-escape) is this one **plus a mid-call escape hatch** (the agent can phone you back when the human asks something off-brief).\n\n```mermaid\nsequenceDiagram\n participant Agent as Your agent (MCP client)\n participant Val as call-my-human (this val)\n participant Vapi as Vapi (gpt-4o + Elliot voice)\n participant Human as 📱 Your human\n\n Agent->>Val: dispatch_call(objective, context, questions)\n Val->>Vapi: POST /call (your Vapi key)\n Vapi->>Human: places the call, runs the whole conversation\n Human-->>Vapi: answers, asks, decides\n Vapi-->>Val: transcripts + end-of-call report (/webhook)\n Agent->>Val: await_report (long-poll until done)\n Val-->>Agent: summary + per-question answers + followups\n```\n\nWhen the call ends, Vapi's **analysis plan** auto-extracts a summary, an answer to each question you asked, and any follow-ups; `await_report` returns it.\n\n## The loop\n\nTwo tools cover the common case:\n\n1. **`dispatch_call`** `{ objective, context?, questions?, number?, first_message?, caller_identity?, max_minutes? }` — the agent hands off the call.\n2. **`await_report`** `{ wait_seconds }` — long-polls; the agent keeps calling it until `done:true`. Returns `{ summary, objective_met, answers[], followups, transcript }`.\n\nOptional supervisor hooks (rarely needed): **`listen`** (eavesdrop on the live call), **`say`** (barge in with a verbatim line), **`get_status`**, **`end_call`**.\n\n## What you need\n\nTwo values from [dashboard.vapi.ai](https://dashboard.vapi.ai):\n\n- your **private key**\n- a **phone number ID** (free Vapi numbers work, US-only — see the reliability note at the bottom)\n\n…plus your **cell number** in E.164 (e.g. `+15551234567`) as the default callee.\n\n## Add it as a connector\n\n### claude.ai (web / desktop)\n\n1. **Settings → Connectors → Add custom connector.**\n2. **Name:** `call-my-human` (anything you like — it's just a label).\n3. **Remote MCP server URL:** paste your connection URL with the three values filled in:\n ```\n https://call-my-human.val.run/mcp?vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\n ```\n4. Leave **OAuth Client ID / Secret** blank, then click **Add**.\n\n> ⚠️ **The `+` in `human_number` must be URL-encoded as `%2B`** (`+15551234567` → `%2B15551234567`).\n\n### Claude Code / Codex (CLI)\n\n```sh\nclaude mcp add --transport http call-my-human \\\n \"https://call-my-human.val.run/mcp?vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\"\n```\n\n`Authorization: Bearer ` is accepted in place of the `vapi_key` param, so CLIs can keep the key out of the URL:\n\n```sh\nclaude mcp add --transport http call-my-human \\\n \"https://call-my-human.val.run/mcp?phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\" \\\n --header \"Authorization: Bearer YOUR_KEY\"\n```\n\nAny MCP client that takes a URL works. The brain runs on your Vapi account's bundled **OpenAI** model (`gpt-4o` default — no extra LLM key needed). Override with `&model=gpt-4o-mini` (faster) or any Vapi-supported model.\n\nThen just tell your agent: *\"if you get stuck, dispatch a call to me and ask.\"*\n\n## Example\n\n> Agent: `dispatch_call({ objective: \"Decide whether to ship the v2 release tonight\", context: \"QA found one minor cosmetic bug, nothing functional. The marketing email is scheduled for 8am.\", questions: [\"Ship tonight or hold for the fix?\", \"If we hold, what's the new target?\"], caller_identity: \"Charlie's AI assistant\" })`\n>\n> …agent calls, talks it through, hangs up…\n>\n> Agent: `await_report({})` → `{ objective_met: true, answers: [{question:\"Ship tonight or hold for the fix?\", answer:\"Ship tonight\"}, {question:\"If we hold, what's the new target?\", answer:\"N/A — shipping\"}], summary: \"...\", followups: \"Wants a Slack ping once it's live.\" }`\n\n## Watching a call live\n\n`dispatch_call` and `get_status` return a `watch_url` (`/?c=&k=`) — a self-refreshing feed of statuses, both-sides transcript, and the final structured report. The `k` is a per-call watch token, so a shared link exposes only that one call, not your whole history.\n\n## Testing with curl\n\n```sh\nBASE=\"https://call-my-human.val.run\"\nQ=\"vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\"\ncurl -X POST \"$BASE/api/call/dispatch?$Q\" -d '{\"objective\":\"Confirm Friday dinner\",\"questions\":[\"Still on for Friday?\",\"What time?\"]}'\ncurl -X POST \"$BASE/api/call/report?$Q\" -d '{\"wait_seconds\":25}' --max-time 35 # repeat until done:true\n```\n\n## Self-host: remix for full privacy 🔒\n\nOn this shared public instance your Vapi key and transcripts pass through a val **you don't own** — the owner could see them in logs/DB. For anything sensitive, **run your own copy.** It's the single biggest privacy lever, and it lets you keep secrets out of the URL entirely.\n\n1. **Remix (fork) the val:** open **[dcm31/call-my-human](https://www.val.town/x/dcm31/call-my-human)** and click **Fork** (top-right). You now own an identical, zero-config copy.\n2. **Put your secrets in environment variables** on your fork (its **Environment Variables** settings) instead of in the URL:\n\n | Env var | Value |\n | --- | --- |\n | `VAPI_KEY` | your Vapi private key |\n | `VAPI_PHONE_NUMBER_ID` | your Vapi phone number ID |\n | `HUMAN_NUMBER` | your cell in E.164, e.g. `+15551234567` |\n | `VAPI_MODEL` *(optional)* | brain model — defaults to `gpt-4o` |\n\n 👉 Set them at: `https://www.val.town/x//call-my-human/environment-variables`\n3. **Connect with a clean URL — no secrets in it.** Grab your fork's HTTP endpoint (shown at the top of `main.ts`) and add `/mcp`:\n ```\n https://.web.val.run/mcp\n ```\n\nPrecedence is **query param / bearer header → env var**, so the public instance (no env set) stays pure BYOK, while your fork can run entirely on env. You can still override the callee per call with the `number` arg or `&human_number=…`.\n\n## Files\n\n- `main.ts` — routes, per-tenant Vapi webhook (transcripts + end-of-call analysis), env/URL credential resolution, landing + live feed\n- `tools.ts` — the tools, shared by MCP and REST; builds the agent's system prompt from your brief\n- `vapi.ts` — Vapi REST client: native-model call creation with the `endCall` tool + structured-data analysis plan\n- `mcp.ts` — minimal stateless streamable-HTTP MCP plumbing (no SDK)\n- `db.ts` — SQLite: tenant-scoped `calls` + append-only `events` (transcripts, report, structured analysis)\n\n## Multi-tenancy & privacy\n\n- Your Vapi key is read per-request and **never stored** (only a hash is logged). It travels TLS-encrypted to this val and on to Vapi, like any BYOK MCP server.\n- Call state, transcripts, and reports **are stored** in this val's SQLite, scoped by a `tenant` = SHA-256(your key) — unguessable without the key, and how Vapi's webhooks find their way back (`/webhook/`, verified via `x-vapi-secret`).\n- **Auto-deleted after 14 days.** A daily cron purges finished calls and their transcripts/events older than `RETENTION_DAYS` (default 14). In-flight calls are never touched — the age cutoff dwarfs the 15-minute max call length. On a remix, set `RETENTION_DAYS` to keep more or less.\n- The only thing kept indefinitely is an **anonymous volume tally**: one timestamp per dispatched call, with no tenant, number, or content. It powers `GET /stats` (calls per day) and nothing else.\n- Want full custody? **Remix it** (see above) — it needs zero configuration.\n\n## Notes\n\n- The agent ends calls itself via Vapi's `endCall` tool once the objective is met; `max_minutes` (default 10) is a hard cap. 30s of mutual silence also ends it.\n- Latency is set by Vapi's STT→LLM→TTS pipeline (~1–2s/turn), not your agent's reasoning loop — that's the whole point vs. the puppet.\n- Post-call analysis (summary + structured data) completes a few seconds *after* the call ends, so `await_report` may return `done:false` once right at hangup — just call it again.\n- The line is clean (no background ambience). Free Vapi numbers are test-grade (occasional dead bridges, a daily outbound limit, narrowband audio). Import a Twilio number into Vapi for reliability.\n- Only call numbers you have consent to call — automated calls without consent violate TCPA/FCC rules.\n","# call-my-human — send an agent to phone your human 📞\n\nA **public remote MCP server**: connect it to any AI agent and it can dispatch an *autonomous* voice agent to phone your human, **run the whole conversation itself**, hang up, and hand back a structured report. Brief it once (`objective` + `context` + `questions`); it talks, ends the call, and you collect the result. The agent has a real LLM brain running **inside Vapi's stack**, so the call is low-latency and natural — you don't drive it turn by turn. **BYOK: no env vars, no server-side keys, no signup.**\n\n> Siblings, same codebase, three control models: [`call-my-human-no-middle-agent`](https://www.val.town/x/dcm31/call-my-human-no-middle-agent) is the *puppet* (you drive every word — full control, high latency); [`dispatch-my-human-escape`](https://www.val.town/x/dcm31/dispatch-my-human-escape) is this one **plus a mid-call escape hatch** (the agent can phone you back when the human asks something off-brief).\n\n```mermaid\nsequenceDiagram\n participant Agent as Your agent (MCP client)\n participant Val as call-my-human (this val)\n participant Vapi as Vapi (gpt-4o + Elliot voice)\n participant Human as 📱 Your human\n\n Agent->>Val: dispatch_call(objective, context, questions)\n Val->>Vapi: POST /call (your Vapi key)\n Vapi->>Human: places the call, runs the whole conversation\n Human-->>Vapi: answers, asks, decides\n Vapi-->>Val: transcripts + end-of-call report (/webhook)\n Agent->>Val: await_report (long-poll until done)\n Val-->>Agent: summary + per-question answers + followups\n```\n\nWhen the call ends, Vapi's **analysis plan** auto-extracts a summary, an answer to each question you asked, and any follow-ups; `await_report` returns it.\n\n## The loop\n\nTwo tools cover the common case:\n\n1. **`dispatch_call`** `{ objective, context?, questions?, number?, first_message?, caller_identity?, max_minutes? }` — the agent hands off the call.\n2. **`await_report`** `{ wait_seconds }` — long-polls; the agent keeps calling it until `done:true`. Returns `{ summary, objective_met, answers[], followups, transcript }`.\n\nOptional supervisor hooks (rarely needed): **`listen`** (eavesdrop on the live call), **`say`** (barge in with a verbatim line), **`get_status`**, **`end_call`**.\n\n## What you need\n\nTwo values from [dashboard.vapi.ai](https://dashboard.vapi.ai):\n\n- your **private key**\n- a **phone number ID** (free Vapi numbers work, US-only — see the reliability note at the bottom)\n\n…plus your **cell number** in E.164 (e.g. `+15551234567`) as the default callee.\n\n## Add it as a connector\n\n### claude.ai (web / desktop)\n\n1. **Settings → Connectors → Add custom connector.**\n2. **Name:** `call-my-human` (anything you like — it's just a label).\n3. **Remote MCP server URL:** paste your connection URL with the three values filled in:\n ```\n https://call-my-human.val.run/mcp?vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\n ```\n4. Leave **OAuth Client ID / Secret** blank, then click **Add**.\n\n> ⚠️ **The `+` in `human_number` must be URL-encoded as `%2B`** (`+15551234567` → `%2B15551234567`).\n\n### Claude Code / Codex (CLI)\n\n```sh\nclaude mcp add --transport http call-my-human \\\n \"https://call-my-human.val.run/mcp?vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\"\n```\n\n`Authorization: Bearer ` is accepted in place of the `vapi_key` param, so CLIs can keep the key out of the URL:\n\n```sh\nclaude mcp add --transport http call-my-human \\\n \"https://call-my-human.val.run/mcp?phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\" \\\n --header \"Authorization: Bearer YOUR_KEY\"\n```\n\nAny MCP client that takes a URL works. The brain runs on your Vapi account's bundled **OpenAI** model (`gpt-4o` default — no extra LLM key needed). Override with `&model=gpt-4o-mini` (faster) or any Vapi-supported model.\n\nThen just tell your agent: *\"if you get stuck, dispatch a call to me and ask.\"*\n\n## Example\n\n> Agent: `dispatch_call({ objective: \"Decide whether to ship the v2 release tonight\", context: \"QA found one minor cosmetic bug, nothing functional. The marketing email is scheduled for 8am.\", questions: [\"Ship tonight or hold for the fix?\", \"If we hold, what's the new target?\"], caller_identity: \"Charlie's AI assistant\" })`\n>\n> …agent calls, talks it through, hangs up…\n>\n> Agent: `await_report({})` → `{ objective_met: true, answers: [{question:\"Ship tonight or hold for the fix?\", answer:\"Ship tonight\"}, {question:\"If we hold, what's the new target?\", answer:\"N/A — shipping\"}], summary: \"...\", followups: \"Wants a Slack ping once it's live.\" }`\n\n## Watching a call live\n\n`dispatch_call` and `get_status` return a `watch_url` (`/?c=&k=`) — a self-refreshing feed of statuses, both-sides transcript, and the final structured report. The `k` is a per-call watch token, so a shared link exposes only that one call, not your whole history.\n\n## Testing with curl\n\n```sh\nBASE=\"https://call-my-human.val.run\"\nQ=\"vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\"\ncurl -X POST \"$BASE/api/call/dispatch?$Q\" -d '{\"objective\":\"Confirm Friday dinner\",\"questions\":[\"Still on for Friday?\",\"What time?\"]}'\ncurl -X POST \"$BASE/api/call/report?$Q\" -d '{\"wait_seconds\":25}' --max-time 35 # repeat until done:true\n```\n\n## Self-host: remix for full privacy 🔒\n\nOn this shared public instance your Vapi key and transcripts pass through a val **you don't own** — the owner could see them in logs/DB. For anything sensitive, **run your own copy.** It's the single biggest privacy lever, and it lets you keep secrets out of the URL entirely.\n\n1. **Remix (fork) the val:** open **[dcm31/call-my-human](https://www.val.town/x/dcm31/call-my-human)** and click **Fork** (top-right). You now own an identical, zero-config copy.\n2. **Put your secrets in environment variables** on your fork (its **Environment Variables** settings) instead of in the URL:\n\n | Env var | Value |\n | --- | --- |\n | `VAPI_KEY` | your Vapi private key |\n | `VAPI_PHONE_NUMBER_ID` | your Vapi phone number ID |\n | `HUMAN_NUMBER` | your cell in E.164, e.g. `+15551234567` |\n | `VAPI_MODEL` *(optional)* | brain model — defaults to `gpt-4o` |\n\n 👉 Set them at: `https://www.val.town/x//call-my-human/environment-variables`\n3. **Connect with a clean URL — no secrets in it.** Grab your fork's HTTP endpoint (shown at the top of `main.ts`) and add `/mcp`:\n ```\n https://.web.val.run/mcp\n ```\n\nPrecedence is **query param / bearer header → env var**, so the public instance (no env set) stays pure BYOK, while your fork can run entirely on env. You can still override the callee per call with the `number` arg or `&human_number=…`.\n\n## Files\n\n- `main.ts` — routes, per-tenant Vapi webhook (transcripts + end-of-call analysis), env/URL credential resolution, landing + live feed\n- `tools.ts` — the tools, shared by MCP and REST; builds the agent's system prompt from your brief\n- `vapi.ts` — Vapi REST client: native-model call creation with the `endCall` tool + structured-data analysis plan\n- `mcp.ts` — minimal stateless streamable-HTTP MCP plumbing (no SDK)\n- `db.ts` — SQLite: tenant-scoped `calls` + append-only `events` (transcripts, report, structured analysis), the anonymous `call_log` tally, and the retention purge\n- `cleanup.ts` — daily interval that purges finished calls/events past `RETENTION_DAYS` (the `call_log` tally is never touched)\n\n## Multi-tenancy & privacy\n\n- Your Vapi key is read per-request and **never stored** (only a hash is logged). It travels TLS-encrypted to this val and on to Vapi, like any BYOK MCP server.\n- Call state, transcripts, and reports **are stored** in this val's SQLite, scoped by a `tenant` = SHA-256(your key) — unguessable without the key, and how Vapi's webhooks find their way back (`/webhook/`, verified via `x-vapi-secret`).\n- **Auto-deleted after 14 days.** A daily cron purges finished calls and their transcripts/events older than `RETENTION_DAYS` (default 14). In-flight calls are never touched — the age cutoff dwarfs the 15-minute max call length. On a remix, set `RETENTION_DAYS` to keep more or less.\n- The only thing kept indefinitely is an **anonymous volume tally**: one timestamp per dispatched call, with no tenant, number, or content. It powers `GET /stats` (calls per day) and nothing else.\n- Want full custody? **Remix it** (see above) — it needs zero configuration.\n\n## Notes\n\n- The agent ends calls itself via Vapi's `endCall` tool once the objective is met; `max_minutes` (default 10) is a hard cap. 30s of mutual silence also ends it.\n- Latency is set by Vapi's STT→LLM→TTS pipeline (~1–2s/turn), not your agent's reasoning loop — that's the whole point vs. the puppet.\n- Post-call analysis (summary + structured data) completes a few seconds *after* the call ends, so `await_report` may return `done:false` once right at hangup — just call it again.\n- The line is clean (no background ambience). Free Vapi numbers are test-grade (occasional dead bridges, a daily outbound limit, narrowband audio). Import a Twilio number into Vapi for reliability.\n- Only call numbers you have consent to call — automated calls without consent violate TCPA/FCC rules.\n","createdAt",["D",1782226793979],"branchVersion","branchId","92e76f90-1592-4578-85fb-c1f07e0fb37e","id","115c4677-1932-4e3d-a81a-e71d5939f255","call-my-human",["D",1781481349040],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","Public remote MCP: your agent phones you. BYO Vapi keys","imageUrl","","9319c175-2593-4f88-ba21-04f2ef7b1c28","pinnedFileId","019ec890-2328-734f-a4be-eb375dc81ac4","pinPreview",true,"pinMigrated","screenshotVersion","likeCount",1,"orgId","d663a026-6449-42f6-9156-1d4a2d64d38c","tursoName","115c467719324e3da81ae71d5939f255","tursoHostname","115c467719324e3da81ae71d5939f255-val-town.aws-us-east-2.turso.io","tursoMeasuredAt",["D",1789509672838],"pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_100":99,"_108":104},"org",{"_100":99,"_101":102,"_103":104,"_18":55,"_105":106,"_107":25},"branch",{"_13":89,"_32":90,"_42":41,"_91":43,"_92":-5,"_93":-5,"_94":-5,"_95":-5,"_96":-5,"_97":44,"_98":99},"isOwned",98,"main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","dcm31","handle","displayName","Charlie","avatar","https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/3e904faa-3764-4629-c475-f561a49b0000/square512?exp=1789556399&sig=79099f5f8ff13fa7ca88ddd17cfb358dd5de01b4572d54739a735b474a8e2cad","tier","pro","canRestrictVals","profileImageUrl","branchName","adoptOrgId",{"_42":43,"_32":44,"_37":261,"_46":-5,"_47":25,"_48":49,"_50":49,"_51":52,"_53":54,"_18":55,"_56":57,"_58":59,"_60":-5,"_61":-5,"_62":63,"_64":65,"_66":67,"_68":69,"_70":262,"_72":-5,"_73":74,"_75":74,"_76":77,"_78":77,"_79":77,"_80":74,"_81":74,"_82":263,"_84":264,"_86":265,"_88":25},"tree",[211,214,215,216,217,218,219,220],"files",[211,212,213,214,215,216,217,218,219,220],"counts",{"_200":186,"_201":63,"_202":63,"_203":204,"_205":206},"contents",[130,131,132,133,134,135,136,137],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},{"_42":193,"_32":194,"_30":195,"_141":-5,"_40":41,"_39":196,"_143":59,"_37":197,"_13":186,"_33":198,"_34":-5,"_24":25,"_148":-5,"_27":199},{"_42":190,"_32":28,"_30":31,"_141":-5,"_40":41,"_39":14,"_143":25,"_37":191,"_13":155,"_33":20,"_34":36,"_24":25,"_148":-5,"_27":192},{"_42":181,"_32":182,"_30":183,"_141":-5,"_40":41,"_39":184,"_143":25,"_37":185,"_13":186,"_33":187,"_34":188,"_24":25,"_148":-5,"_27":189},{"_42":174,"_32":175,"_30":176,"_141":-5,"_40":41,"_39":177,"_143":25,"_37":178,"_13":163,"_33":146,"_34":179,"_24":25,"_148":-5,"_27":180},{"_42":166,"_32":167,"_30":57,"_141":-5,"_40":41,"_39":168,"_143":25,"_37":169,"_13":170,"_33":171,"_34":172,"_24":25,"_148":44,"_27":173},{"_42":158,"_32":159,"_30":160,"_141":-5,"_40":41,"_39":161,"_143":59,"_37":162,"_13":163,"_33":146,"_34":164,"_24":25,"_148":-5,"_27":165},{"_42":150,"_32":151,"_30":152,"_141":-5,"_40":41,"_39":153,"_143":25,"_37":154,"_13":155,"_33":146,"_34":156,"_24":25,"_148":-5,"_27":157},{"_42":138,"_32":139,"_30":140,"_141":-5,"_40":41,"_39":142,"_143":59,"_37":144,"_13":145,"_33":146,"_34":147,"_24":25,"_148":-5,"_27":149},"dbbdc6ea-e246-4b06-8747-ff2c02c3e156","vapi.ts","019ec890-2328-734f-a4be-f6a0f6422d3f","parentId",48,"latest",["D",1782163327615],11,"script","// Thin Vapi REST client — BYOK: every function takes the caller's own key.\n// Nothing is read from env; this is a public multi-tenant server.\n//\n// call-my-human DOES use Vapi's own LLM brain (native model), unlike the\n// puppet sibling call-my-human-no-middle-agent whose custom-llm always returned \"\". The agent\n// runs the whole conversation itself for low latency, then Vapi's analysis\n// produces a structured report we hand back to the dispatcher.\n\nconst VAPI_BASE = \"https://api.vapi.ai\";\n\nexport interface DispatchOpts {\n vapiKey: string;\n phoneNumberId: string;\n number: string;\n firstMessage: string;\n systemPrompt: string;\n questions: string[];\n model: string; // e.g. \"gpt-4o\"\n maxDurationSeconds: number;\n origin: string; // this val's https origin\n tenant: string; // hash of the vapi key — routes webhooks back to the right feed (in the URL path)\n webhookSecret: string; // per-call random secret Vapi echoes back as x-vapi-secret (auth, not identity)\n}\n\n// Structured report schema: objective + one answer per required question + followups.\n// The required questions are baked into the schema descriptions (not a separate\n// prompt) so the extractor always sees them — Vapi feeds it the transcript by default.\nfunction structuredSchema(questions: string[]) {\n const qList = questions.length\n ? \" The agent was told to get answers to these specific questions: \" +\n questions.map((q, i) => `(${i + 1}) ${q}`).join(\"; \") +\n \". Provide one entry per question, in that order.\"\n : \"\";\n return {\n type: \"object\",\n properties: {\n objective_met: {\n type: \"boolean\",\n description: \"True if the call accomplished its stated objective.\",\n },\n answers: {\n type: \"array\",\n description:\n \"One entry for each question the agent was asked to get answered, with the human's answer.\" + qList,\n items: {\n type: \"object\",\n properties: {\n question: { type: \"string\", description: \"The question the agent needed answered.\" },\n answer: {\n type: \"string\",\n description: \"The human's answer, paraphrased or verbatim. Use 'not answered' if never resolved.\",\n },\n },\n required: [\"question\", \"answer\"],\n },\n },\n followups: {\n type: \"string\",\n description:\n \"Anything the human raised, requested, or asked that the dispatcher should act on after the call. Empty string if none.\",\n },\n },\n required: [\"objective_met\", \"answers\"],\n };\n}\n\nexport async function createDispatchCall(opts: DispatchOpts) {\n const body = {\n phoneNumberId: opts.phoneNumberId,\n customer: { number: opts.number },\n assistant: {\n firstMessage: opts.firstMessage,\n // A REAL brain this time: Vapi runs the LLM in-stack (lowest latency, no\n // round-trip back to the dispatcher). The endCall tool lets it hang up\n // itself once the objective is met.\n model: {\n provider: \"openai\",\n model: opts.model,\n temperature: 0.6,\n messages: [{ role: \"system\", content: opts.systemPrompt }],\n tools: [{ type: \"endCall\" }],\n },\n // No transcriber override — use Vapi's default (what templates/vapi-phone-call\n // uses). The explicit deepgram nova-2 override added latency and produced\n // duplicated/cumulative transcript lines. Native models get a default STT.\n // Vapi-native voice — what the dashboard \"test call\" uses, tuned for telephony\n // and bundled (no extra provider key). Much better on a phone than openai/shimmer.\n voice: { provider: \"vapi\", voiceId: \"Elliot\", version: 2 },\n // No ambient soundscape — Vapi defaults to an \"office\" background hum that\n // sounds odd on a 1:1 agent call. \"off\" = clean line.\n backgroundSound: \"off\",\n server: { url: `${opts.origin}/webhook/${opts.tenant}`, secret: opts.webhookSecret },\n serverMessages: [\"transcript\", \"status-update\", \"end-of-call-report\", \"conversation-update\"],\n // The human pauses to think; the agent answers immediately, so dead air\n // means a genuinely dropped/empty call — 30s is a safe cutoff.\n silenceTimeoutSeconds: 30,\n maxDurationSeconds: opts.maxDurationSeconds,\n // Auto-summarize + extract the structured report when the call ends.\n // Schema only — Vapi's default extractor prompt already receives the\n // transcript; a custom {{transcript}} message silently broke extraction.\n analysisPlan: {\n structuredDataPlan: {\n enabled: true,\n schema: structuredSchema(opts.questions),\n },\n },\n },\n };\n\n console.log(\"vapi: dispatching call →\", opts.number, \"tenant\", opts.tenant, \"model\", opts.model);\n const res = await fetch(`${VAPI_BASE}/call`, {\n method: \"POST\",\n headers: { Authorization: `Bearer ${opts.vapiKey}`, \"Content-Type\": \"application/json\" },\n body: JSON.stringify(body),\n });\n const data = await res.json();\n if (!res.ok) {\n console.log(\"vapi: dispatch FAILED\", res.status, JSON.stringify(data).slice(0, 500));\n throw new Error(`Vapi ${res.status}: ${JSON.stringify(data)}`);\n }\n console.log(\"vapi: call created\", data.id, \"controlUrl:\", data.monitor?.controlUrl ? \"present\" : \"MISSING\");\n return data as { id: string; status?: string; monitor?: { controlUrl?: string; listenUrl?: string } };\n}\n\nexport async function control(controlUrl: string, payload: Record) {\n const res = await fetch(controlUrl, {\n method: \"POST\",\n headers: { \"Content-Type\": \"application/json\" },\n body: JSON.stringify(payload),\n });\n const text = await res.text();\n console.log(\"vapi control:\", payload.type, \"→\", res.status, text.slice(0, 200));\n return { status: res.status, body: text };\n}\n\n// Optional supervisor barge-in: inject a verbatim line over the live call.\n// Docs use `content`; the SDK's SayMessage type uses `message`. Try both.\nexport async function sayOnCall(controlUrl: string, text: string, endCallAfterSpoken = false) {\n const base: Record = { type: \"say\", content: text };\n if (endCallAfterSpoken) base.endCallAfterSpoken = true;\n let r = await control(controlUrl, base);\n if (r.status >= 400) {\n const alt: Record = { type: \"say\", message: text };\n if (endCallAfterSpoken) alt.endCallAfterSpoken = true;\n r = await control(controlUrl, alt);\n }\n return r;\n}\n\nexport async function endCallOnVapi(controlUrl: string) {\n return control(controlUrl, { type: \"end-call\" });\n}\n\nexport async function getCallLive(vapiKey: string, callId: string) {\n const res = await fetch(`${VAPI_BASE}/call/${callId}`, {\n headers: { Authorization: `Bearer ${vapiKey}` },\n });\n const data = await res.json();\n if (!res.ok) throw new Error(`Vapi GET /call ${res.status}: ${JSON.stringify(data).slice(0, 300)}`);\n return data;\n}\n","label",[139],"63243062-3259-4fe5-81bd-c1c194595883","tools.ts","019ec890-2328-734f-a4be-f0d464ec163a",74,["D",1782226727210],10,"// The tools, defined once and exposed twice: over MCP (/mcp) and REST (/api/call/*).\n// All config arrives per-request (BYOK) via ctx — built in main.ts from the\n// connection URL's query params. Nothing comes from env.\n//\n// Unlike the puppet sibling (call-my-human-no-middle-agent), here the Vapi agent has a REAL\n// brain and runs the whole conversation itself. Primary loop:\n// dispatch_call → hand off a brief + objective + questions\n// await_report → long-poll until it hangs up, get the structured result\n// listen / say are optional supervisor hooks (eavesdrop / barge in).\n\nimport * as db from \"./db.ts\";\nimport { createDispatchCall, endCallOnVapi, getCallLive, sayOnCall } from \"./vapi.ts\";\n\nexport interface Ctx {\n origin: string;\n vapiKey: string | null;\n phoneNumberId: string | null;\n humanNumber: string | null;\n model: string; // LLM the agent's brain runs on (default gpt-4o)\n tenant: string | null; // short hash of vapiKey\n}\n\nconst CONNECT_HELP =\n \"Missing Vapi credentials. Put them in your MCP connection URL: /mcp?vapi_key=&phone_number_id=&human_number=%2B15551234567 (the + must be URL-encoded as %2B). Get the key and phone number at dashboard.vapi.ai.\";\n\nfunction creds(ctx: Ctx): { vapiKey: string; tenant: string } {\n if (!ctx.vapiKey || !ctx.tenant) throw new Error(CONNECT_HELP);\n return { vapiKey: ctx.vapiKey, tenant: ctx.tenant };\n}\n\nconst ACTIVE = db.ACTIVE_STATUSES;\n\nasync function requireActiveCall(tenant: string): Promise {\n const call = await db.latestCall(tenant);\n if (!call) throw new Error(\"No call yet — use dispatch_call first.\");\n if (!ACTIVE.includes(call.status)) {\n throw new Error(\n `Latest call ${call.id} is '${call.status}'${call.ended_reason ? ` (${call.ended_reason})` : \"\"} — start a new one with dispatch_call.`,\n );\n }\n return call;\n}\n\nconst fmt = (e: db.EventRow) => ({\n t: e.created_at.slice(11, 19),\n type: e.type,\n role: e.role,\n content: e.content,\n});\n\n// Build the agent's system prompt from the dispatch brief. This IS the brain —\n// it carries the whole call, so be explicit about who it is and what \"done\" means.\nfunction buildSystemPrompt(opts: {\n objective: string;\n context: string;\n questions: string[];\n callerIdentity: string;\n maxMinutes: number;\n}): string {\n const qBlock = opts.questions.length\n ? \"Before you hang up, get a clear answer to EACH of these:\\n\" +\n opts.questions.map((q, i) => ` ${i + 1}. ${q}`).join(\"\\n\")\n : \"There are no specific required questions beyond the objective above.\";\n const ctxBlock = opts.context.trim()\n ? `Context you've been briefed with (use it; don't read it aloud verbatim):\\n${opts.context.trim()}`\n : \"You have not been given extra context beyond the objective.\";\n\n return `You are ${opts.callerIdentity}. You are placing an OUTBOUND phone call to a person on behalf of your operator (the AI assistant / system that dispatched you), who is busy and needs something from them. You called THEM — you are not their assistant waiting for requests.\n\nYOUR OBJECTIVE:\n${opts.objective.trim()}\n\n${ctxBlock}\n\n${qBlock}\n\nHOW TO TALK:\n- Sound like a real person on the phone: short, natural sentences, one idea at a time. No monologues.\n- Open by quickly saying who you are and why you're calling, then get to the point.\n- Listen and react to what they actually say. Don't plow through a script.\n- If they ask something you weren't briefed on, be honest: say you'll relay it back to your operator rather than guessing or making things up.\n- If it's a bad time, offer to have your operator follow up later, then wrap up.\n\nWHEN YOU'RE DONE:\n- NEVER end the call while the human is still talking or mid-thought. Wait for a clear, complete answer and a natural pause.\n- Only after you have answers to ALL the required questions (or they say they have to go) should you wrap up. First ask \"anything else before I let you go?\" — if they're still talking, keep listening; do not hang up.\n- Then briefly confirm what you heard, thank them, and use the endCall function.\n- Keep the whole call under ${opts.maxMinutes} minutes, but getting complete answers matters more than speed — don't rush them off.`;\n}\n\nexport const TOOLS = [\n {\n name: \"dispatch_call\",\n description:\n \"Dispatch a voice agent to phone your human and handle a conversation autonomously, then hang up. UNLIKE a puppet, this agent has a real brain: you hand it a brief once and it carries the whole call itself with low latency — you do NOT drive it turn by turn. Give it a clear `objective`, any `context` it needs, and the specific `questions` you want answered. After it returns, call `await_report` to block until the call ends and collect the structured result (summary + per-question answers + followups). For most check-ins this is the only tool you need besides await_report.\",\n inputSchema: {\n type: \"object\",\n properties: {\n objective: {\n type: \"string\",\n description: \"The goal of the call in plain language, e.g. 'Confirm whether we should ship the release tonight or wait for QA.'\",\n },\n context: {\n type: \"string\",\n description: \"Everything the agent should know to handle the call well — the situation, relevant facts, constraints. Dump freely; it won't read it aloud.\",\n },\n questions: {\n type: \"array\",\n items: { type: \"string\" },\n description: \"Specific questions you need answered. Each becomes a field in the returned report.\",\n },\n number: {\n type: \"string\",\n description: \"E.164 number to dial, e.g. +15551234567. Defaults to the human_number from your connection URL.\",\n },\n first_message: {\n type: \"string\",\n description: \"The exact opening line the agent speaks on pickup. Optional — a sensible default is used.\",\n },\n caller_identity: {\n type: \"string\",\n description: \"Who the agent should present as, e.g. 'Charlie's AI assistant'. Optional.\",\n },\n max_minutes: {\n type: \"number\",\n description: \"Hard cap on call length in minutes (default 10).\",\n },\n },\n required: [\"objective\"],\n },\n run: async (args: Record, ctx: Ctx) => {\n const { vapiKey, tenant } = creds(ctx);\n const phoneNumberId = ctx.phoneNumberId || String(args.phone_number_id || \"\");\n if (!phoneNumberId) {\n throw new Error(\"phone_number_id missing — add &phone_number_id= to your connection URL.\");\n }\n const prev = await db.latestCall(tenant);\n if (prev && ACTIVE.includes(prev.status)) {\n let stillActive = false;\n try {\n const live = await getCallLive(vapiKey, prev.id);\n if (live.status && live.status !== prev.status) {\n await db.setCallStatus(prev.id, live.status, live.endedReason ?? null);\n }\n stillActive = ACTIVE.includes(live.status);\n } catch (e) {\n console.log(\"dispatch_call: live status check failed, allowing new call:\", String(e).slice(0, 120));\n }\n if (stillActive) throw new Error(`Call ${prev.id} is still active — end_call it first.`);\n }\n const number = String(args.number || ctx.humanNumber || \"\");\n if (!number) {\n throw new Error(\n \"No number to dial — pass `number`, or add &human_number=%2B15551234567 to your connection URL.\",\n );\n }\n const objective = String(args.objective || \"\").trim();\n if (!objective) throw new Error(\"`objective` is required — tell the agent what the call is for.\");\n const context = String(args.context || \"\");\n const questions = Array.isArray(args.questions) ? args.questions.map(String).filter(Boolean) : [];\n const callerIdentity = String(args.caller_identity || \"an AI voice assistant\");\n const maxMinutes = Math.min(Math.max(Number(args.max_minutes ?? 10), 1), 15);\n const firstMessage = String(\n args.first_message ||\n \"Hi! This is your AI assistant calling — I've got a quick thing to run by you, do you have a sec?\",\n );\n const systemPrompt = buildSystemPrompt({ objective, context, questions, callerIdentity, maxMinutes });\n\n // Per-call random secrets: the webhook secret authenticates Vapi's inbound\n // events; the watch token is a single-call bearer capability for the live\n // feed. Neither is the tenant — so leaking one watch link exposes one call.\n const webhookSecret = db.randomToken();\n const watchToken = db.randomToken();\n\n const call = await createDispatchCall({\n vapiKey,\n phoneNumberId,\n number,\n firstMessage,\n systemPrompt,\n questions,\n model: ctx.model,\n maxDurationSeconds: maxMinutes * 60,\n origin: ctx.origin,\n tenant,\n webhookSecret,\n });\n const controlUrl = call.monitor?.controlUrl ?? \"\";\n await db.insertCall(tenant, call.id, controlUrl, number, watchToken, webhookSecret);\n await db.logCall(); // anonymous calls/day tally — survives the retention purge\n await db.addEvent(tenant, call.id, \"status\", \"system\", `dispatching ${number} · objective: ${objective.slice(0, 120)}`);\n return {\n call_id: call.id,\n status: call.status ?? \"queued\",\n dialing: number,\n questions_tracked: questions.length,\n watch_url: `${ctx.origin}/?c=${call.id}&k=${watchToken}`,\n next:\n \"Call `await_report` now (it long-polls) and keep calling it until it returns done:true with the summary + answers. To eavesdrop or barge in mid-call, use `listen` / `say`.\",\n };\n },\n },\n {\n name: \"await_report\",\n description:\n \"Block until the dispatched call finishes, then return the structured outcome: a summary, an answer to each question you asked, anything the human raised for follow-up, and the full transcript. Long-polls up to wait_seconds — keep calling it (it's cheap) until you get done:true. While done:false the call is still in progress or the post-call analysis is still being written.\",\n inputSchema: {\n type: \"object\",\n properties: {\n wait_seconds: { type: \"number\", description: \"Max seconds to long-poll this call (1-25, default 20).\" },\n },\n },\n run: async (args: Record, ctx: Ctx) => {\n const { vapiKey, tenant } = creds(ctx);\n const call = await db.latestCall(tenant);\n if (!call) throw new Error(\"No call yet — use dispatch_call first.\");\n const wait = Math.min(Math.max(Number(args.wait_seconds ?? 20), 1), 25);\n const deadline = Date.now() + wait * 1000;\n\n const buildReport = async (status: string, endedReason: string | null) => {\n const analysis = await db.latestEventOfType(call.id, \"analysis\");\n const summaryEv = await db.latestEventOfType(call.id, \"report\");\n let structured: Record | null = null;\n if (analysis?.content) {\n try {\n structured = JSON.parse(analysis.content);\n } catch { /* leave null */ }\n }\n const transcript = (await db.transcriptLines(call.id)).map((e) => ({ role: e.role, text: e.content }));\n return {\n done: true,\n call_status: status,\n ended_reason: endedReason,\n summary: summaryEv?.content ?? null,\n objective_met: structured?.objective_met ?? null,\n answers: structured?.answers ?? null,\n followups: structured?.followups ?? null,\n transcript,\n ...(structured ? {} : { note: \"Call ended but structured analysis isn't ready yet — call await_report once more in a few seconds.\" }),\n };\n };\n\n while (true) {\n let current = await db.getCall(call.id);\n const terminal = current && !ACTIVE.includes(current.status);\n // Done = call ended AND the post-call analysis has landed (or we've waited it out).\n if (terminal) {\n const analysis = await db.latestEventOfType(call.id, \"analysis\");\n if (analysis) return await buildReport(current!.status, current!.ended_reason);\n }\n if (Date.now() > deadline) {\n // Sync from Vapi in case a webhook was missed.\n try {\n const live = await getCallLive(vapiKey, call.id);\n if (live?.status && live.status !== current?.status) {\n await db.setCallStatus(call.id, String(live.status), (live.endedReason as string) ?? null);\n current = await db.getCall(call.id);\n }\n } catch (e) {\n console.log(\"await_report: live status check failed:\", String(e).slice(0, 120));\n }\n const stillTerminal = current && !ACTIVE.includes(current.status);\n if (stillTerminal) return await buildReport(current!.status, current!.ended_reason);\n return {\n done: false,\n call_status: current?.status ?? \"unknown\",\n ...(current?.live_partial ? { hearing_now: current.live_partial } : {}),\n note: \"Call still in progress — call await_report again to keep waiting.\",\n };\n }\n await new Promise((r) => setTimeout(r, 1000));\n }\n },\n },\n {\n name: \"listen\",\n description:\n \"Optional: eavesdrop on a live dispatched call. Long-polls for new events (the human's speech, the agent's lines, status changes) up to wait_seconds. Use this if you want to watch the conversation unfold in realtime; you do NOT need it for the normal dispatch→await_report flow.\",\n inputSchema: {\n type: \"object\",\n properties: {\n wait_seconds: { type: \"number\", description: \"Max seconds to wait (1-25, default 20).\" },\n },\n },\n run: async (args: Record, ctx: Ctx) => {\n const { vapiKey, tenant } = creds(ctx);\n const call = await db.latestCall(tenant);\n if (!call) throw new Error(\"No call yet — use dispatch_call first.\");\n const wait = Math.min(Math.max(Number(args.wait_seconds ?? 20), 1), 25);\n const deadline = Date.now() + wait * 1000;\n let cursor = call.cursor;\n while (true) {\n const events = await db.eventsAfter(call.id, cursor);\n const maxId = events.length ? events[events.length - 1].id : cursor;\n if (events.length) {\n await db.setCursor(call.id, maxId);\n const current = await db.getCall(call.id);\n const over = current && !ACTIVE.includes(current.status);\n return {\n events: events.map(fmt),\n call_status: current?.status ?? call.status,\n ...(over ? { note: `Call is over (${current?.ended_reason ?? current?.status}). Use await_report for the result.` } : {}),\n };\n }\n const current = await db.getCall(call.id);\n if (current && !ACTIVE.includes(current.status)) {\n return {\n events: [],\n call_status: current.status,\n note: `Call is over (${current.ended_reason ?? current.status}). Use await_report for the result.`,\n };\n }\n if (Date.now() > deadline) {\n try {\n const live = await getCallLive(vapiKey, call.id);\n if (live?.status && live.status !== current?.status) {\n await db.setCallStatus(call.id, String(live.status), (live.endedReason as string) ?? null);\n }\n } catch (e) {\n console.log(\"listen: live status check failed:\", String(e).slice(0, 120));\n }\n const synced = await db.getCall(call.id);\n if (synced && !ACTIVE.includes(synced.status)) {\n return {\n events: [],\n call_status: synced.status,\n note: `Call is over (${synced.ended_reason ?? synced.status}). Use await_report for the result.`,\n };\n }\n return {\n events: [],\n call_status: synced?.status ?? \"unknown\",\n ...(synced?.live_partial ? { hearing_now: synced.live_partial } : {}),\n note: \"Nothing new yet — the agent is handling it. listen again, or just await_report.\",\n };\n }\n await new Promise((r) => setTimeout(r, 800));\n }\n },\n },\n {\n name: \"say\",\n description:\n \"Optional supervisor barge-in: speak a verbatim line over the live call yourself (TTS), e.g. to correct the agent or relay a fact it didn't have. Normally unnecessary — the agent drives the call. Set end_call_after true to hang up after the line plays.\",\n inputSchema: {\n type: \"object\",\n properties: {\n text: { type: \"string\", description: \"What to say, verbatim.\" },\n end_call_after: { type: \"boolean\", description: \"Hang up after this line finishes playing.\" },\n },\n required: [\"text\"],\n },\n run: async (args: Record, ctx: Ctx) => {\n const { tenant } = creds(ctx);\n const call = await requireActiveCall(tenant);\n if (!call.control_url) throw new Error(\"No controlUrl stored for this call — cannot inject speech.\");\n const r = await sayOnCall(call.control_url, String(args.text), Boolean(args.end_call_after));\n await db.addEvent(tenant, call.id, \"say\", \"assistant\", String(args.text));\n if (r.status >= 400) throw new Error(`Vapi control rejected say (${r.status}): ${r.body.slice(0, 200)}`);\n return { ok: true, vapi_status: r.status, next: \"Use listen to hear their response, or await_report for the final result.\" };\n },\n },\n {\n name: \"get_status\",\n description:\n \"Snapshot of your latest call: live status straight from Vapi plus the recent event feed. Does not consume listen's cursor. Includes a watch_url where you can see the live transcript.\",\n inputSchema: { type: \"object\", properties: {} },\n run: async (_args: Record, ctx: Ctx) => {\n const { vapiKey, tenant } = creds(ctx);\n const call = await db.latestCall(tenant);\n if (!call) return { note: \"No calls yet.\" };\n let live: Record | null = null;\n try {\n live = await getCallLive(vapiKey, call.id);\n if (live?.status && live.status !== call.status) {\n await db.setCallStatus(call.id, String(live.status), (live.endedReason as string) ?? null);\n }\n } catch (e) {\n live = { error: String(e) };\n }\n const events = await db.recentEvents(call.id, 15);\n return {\n call_id: call.id,\n number: call.number,\n db_status: call.status,\n vapi_status: live?.status ?? null,\n ended_reason: live?.endedReason ?? call.ended_reason ?? null,\n started_at: call.created_at,\n watch_url: call.watch_token ? `${ctx.origin}/?c=${call.id}&k=${call.watch_token}` : null,\n recent_events: events.map(fmt),\n };\n },\n },\n {\n name: \"end_call\",\n description: \"Hang up the live call immediately (e.g. to abort a dispatched call). The agent normally ends calls itself.\",\n inputSchema: { type: \"object\", properties: {} },\n run: async (_args: Record, ctx: Ctx) => {\n const { tenant } = creds(ctx);\n const call = await requireActiveCall(tenant);\n const r = await endCallOnVapi(call.control_url);\n await db.addEvent(tenant, call.id, \"status\", \"system\", \"end-call requested\");\n return { ok: r.status < 400, vapi_status: r.status };\n },\n },\n];\n\nexport async function runTool(name: string, args: Record, ctx: Ctx) {\n const tool = TOOLS.find((t) => t.name === name);\n if (!tool) throw new Error(`Unknown tool: ${name}`);\n console.log(\"tool:\", name, \"tenant:\", ctx.tenant ?? \"none\", JSON.stringify(args ?? {}).slice(0, 200));\n return await tool.run(args ?? {}, ctx);\n}\n",[151],"fe535f0c-3f36-4440-910c-37f10dfec87f","mcp.ts","019ec890-2328-734f-a4be-ef33318fbd23",45,["D",1782155923638],5,"// Minimal remote MCP server: stateless streamable HTTP, plain JSON responses,\n// no SDK. Handles initialize / ping / tools/list / tools/call; notifications get 202.\n\nimport { type Ctx, runTool, TOOLS } from \"./tools.ts\";\n\nconst KNOWN_VERSIONS = [\"2024-11-05\", \"2025-03-26\", \"2025-06-18\"];\n\nconst SERVER_INSTRUCTIONS =\n `call-my-human: dispatch an autonomous voice agent to phone your human, run the whole call itself, and report back.\nThis agent has a REAL brain — you do NOT drive it word by word. You brief it once and it carries the conversation.\nPrimary loop:\n 1. dispatch_call({ objective, context?, questions?, number?, first_message?, max_minutes? }) — hand off the call.\n 2. await_report({ wait_seconds }) — long-polls; keep calling it until it returns done:true with { summary, objective_met, answers[], followups, transcript }.\nGive a crisp objective, dump any context the agent needs, and list the exact questions you want answered (each comes back as a field).\nOptional supervisor hooks: listen (eavesdrop on the live call) and say (barge in with a verbatim line). get_status / end_call as needed.\nSetup: your connection URL must carry vapi_key and phone_number_id (and ideally human_number, E.164 with + encoded as %2B) from dashboard.vapi.ai. The agent's brain runs on your Vapi account's OpenAI model (gpt-4o default; &model= to override).`;\n\nconst rpcResult = (id: unknown, result: unknown) => ({ jsonrpc: \"2.0\", id, result });\nconst rpcError = (id: unknown, code: number, message: string) => ({ jsonrpc: \"2.0\", id, error: { code, message } });\n\nasync function handleOne(msg: any, ctx: Ctx) {\n if (!msg || typeof msg !== \"object\" || typeof msg.method !== \"string\") {\n return rpcError(msg?.id ?? null, -32600, \"invalid request\");\n }\n if (msg.id === undefined) {\n console.log(\"mcp notification:\", msg.method);\n return null; // notifications get no body\n }\n console.log(\"mcp request:\", msg.method, \"tenant:\", ctx.tenant ?? \"none\");\n switch (msg.method) {\n case \"initialize\": {\n const v = msg.params?.protocolVersion;\n return rpcResult(msg.id, {\n protocolVersion: KNOWN_VERSIONS.includes(v) ? v : \"2025-06-18\",\n capabilities: { tools: { listChanged: false } },\n serverInfo: {\n name: \"call-my-human\",\n version: \"1.0.0\",\n // MCP SEP-973 (spec 2025-11-25): visual identifier for this server. Clients\n // don't all render this yet (claude.ai connector-icon support is an open\n // request), but the ones that do will pick up the 📞 automatically.\n icons: [{ src: \"https://call-my-human.val.run/favicon.svg\", mimeType: \"image/svg+xml\" }],\n },\n instructions: SERVER_INSTRUCTIONS,\n });\n }\n case \"ping\":\n return rpcResult(msg.id, {});\n case \"tools/list\":\n return rpcResult(msg.id, {\n tools: TOOLS.map(({ name, description, inputSchema }) => ({ name, description, inputSchema })),\n });\n case \"tools/call\": {\n try {\n const out = await runTool(msg.params?.name, msg.params?.arguments ?? {}, ctx);\n const text = typeof out === \"string\" ? out : JSON.stringify(out, null, 2);\n return rpcResult(msg.id, { content: [{ type: \"text\", text }], isError: false });\n } catch (e) {\n return rpcResult(msg.id, { content: [{ type: \"text\", text: String(e) }], isError: true });\n }\n }\n default:\n return rpcError(msg.id, -32601, `method not found: ${msg.method}`);\n }\n}\n\nexport async function handleMcp(req: Request, ctx: Ctx): Promise {\n if (req.method !== \"POST\") {\n return new Response(\"call-my-human: POST JSON-RPC here (stateless streamable HTTP MCP). GET / for setup.\", {\n status: 405,\n });\n }\n let body: unknown;\n try {\n body = await req.json();\n } catch {\n return Response.json(rpcError(null, -32700, \"parse error\"), { status: 400 });\n }\n const msgs = Array.isArray(body) ? body : [body];\n const replies: unknown[] = [];\n for (const m of msgs) {\n const r = await handleOne(m, ctx);\n if (r) replies.push(r);\n }\n if (!replies.length) return new Response(null, { status: 202 });\n return Response.json(Array.isArray(body) ? replies : replies[0]);\n}\n",[159],"8c0329fc-7d7a-4507-aa65-018683d38d74","main.ts",76,["D",1782226751300],32,"http","// call-my-human — a PUBLIC remote MCP server that dispatches an autonomous\n// voice agent to phone your human, handle the conversation itself, and report back.\n// Unlike the puppet sibling (call-my-human-no-middle-agent), the Vapi agent here has a REAL brain:\n// it runs the whole call in-stack (low latency, no per-turn round-trip), then Vapi's\n// post-call analysis returns a structured report.\n// BYOK: connecting users pass their own Vapi creds in the connection URL.\n// No env vars, no server-side keys, no signup.\n//\n// /mcp?vapi_key=…&phone_number_id=…&human_number=%2B1555… ← connect here\n// /api/call/:action same tools as REST (dispatch, report, listen, say, status, end)\n// /webhook/:tenant Vapi server events → SQLite (transcripts + end-of-call analysis)\n// / plain GET = setup instructions; /?c=&k= = live transcript feed\n//\n// A tenant is a short hash of the caller's Vapi key — it scopes call state per user\n// and routes Vapi's webhooks back to the right feed. Keys are never stored.\n\nimport { Hono } from \"npm:hono\";\nimport { parseVal, readFile } from \"https://esm.town/v/std/utils/index.ts\";\nimport * as db from \"./db.ts\";\nimport { handleMcp } from \"./mcp.ts\";\nimport { type Ctx, runTool } from \"./tools.ts\";\n\nconst app = new Hono();\napp.onError((err) => Promise.reject(err));\n\nasync function tenantOf(key: string): Promise {\n // Full SHA-256 (not truncated). The tenant is internal-only — it scopes a user's\n // calls and routes webhooks; it's never a capability token and never in a URL.\n const d = await crypto.subtle.digest(\"SHA-256\", new TextEncoder().encode(key));\n return [...new Uint8Array(d)].map((b) => b.toString(16).padStart(2, \"0\")).join(\"\");\n}\n\nasync function buildCtx(c: any): Promise {\n const q = (k: string) => c.req.query(k) || null;\n const bearer = (c.req.header(\"authorization\") ?? \"\").replace(/^Bearer\\s+/i, \"\") || null;\n // Precedence: per-request (query param / bearer header) wins, then env vars.\n // Env only matters on a REMIXED, single-user instance — set VAPI_KEY etc. and\n // connect with a bare /mcp URL (no secrets in the URL). The public shared server\n // has no env set, so it stays pure BYOK.\n const vapiKey = q(\"vapi_key\") ?? bearer ?? Deno.env.get(\"VAPI_KEY\") ?? null;\n return {\n origin: new URL(c.req.url).origin,\n vapiKey,\n phoneNumberId: q(\"phone_number_id\") ?? Deno.env.get(\"VAPI_PHONE_NUMBER_ID\") ?? null,\n humanNumber: q(\"human_number\") ?? Deno.env.get(\"HUMAN_NUMBER\") ?? null,\n model: q(\"model\") || Deno.env.get(\"VAPI_MODEL\") || \"gpt-4o\", // agent's brain; &model= or env to override\n tenant: vapiKey ? await tenantOf(vapiKey) : null,\n };\n}\n\n// ---------- remote MCP ----------\napp.all(\"/mcp\", async (c) => handleMcp(c.req.raw, await buildCtx(c)));\n\n// ---------- REST mirror of the tools ----------\nconst REST: Record = {\n dispatch: \"dispatch_call\",\n report: \"await_report\",\n listen: \"listen\",\n say: \"say\",\n status: \"get_status\",\n end: \"end_call\",\n};\n\napp.all(\"/api/call/:action\", async (c) => {\n const tool = REST[c.req.param(\"action\")];\n if (!tool) return c.json({ error: \"unknown action — use dispatch | report | listen | say | status | end\" }, 404);\n let args: Record = {};\n if (c.req.method === \"POST\") {\n try {\n args = await c.req.json();\n } catch { /* empty body is fine */ }\n }\n try {\n return c.json({ ok: true, result: await runTool(tool, args, await buildCtx(c)) });\n } catch (e) {\n return c.json({ ok: false, error: String(e) }, 400);\n }\n});\n\n// State for the live feed page. Scoped to ONE call by its id + per-call watch token\n// (a 128-bit bearer capability from the watch_url) — not the tenant. Sensitive\n// fields (control_url, secrets, tenant) are stripped from the response.\napp.get(\"/api/state\", async (c) => {\n const callId = c.req.query(\"c\");\n const token = c.req.query(\"k\");\n if (!callId || !token) {\n return c.json({ error: \"pass ?c=&k= from the watch_url\" }, 400);\n }\n const call = await db.getCall(callId);\n if (!call || !call.watch_token || call.watch_token !== token) {\n return c.json({ error: \"call not found, or bad watch token\" }, 403);\n }\n const events = await db.recentEvents(call.id, 120);\n const { control_url: _cu, webhook_secret: _ws, watch_token: _wt, tenant: _tn, ...safe } = call;\n return c.json({ call: safe, events });\n});\n\n// Anonymous volume stats: calls/day across the whole server. No tenant, number,\n// or content — just counts. Aggregate-only, so it's safe to leave ungated.\napp.get(\"/stats\", async (c) => {\n const perDay = await db.callsPerDay(90);\n const total = perDay.reduce((s, d) => s + d.count, 0);\n return c.json({ total_logged: total, per_day: perDay });\n});\n\n// ---------- Vapi webhook (per tenant) ----------\napp.post(\"/webhook/:tenant\", async (c) => {\n const routeTenant = c.req.param(\"tenant\");\n const payload = await c.req.json().catch(() => ({}));\n const m = payload?.message ?? {};\n // Resolve the call this webhook is about, then verify against ITS per-call\n // secret (not the tenant). The tenant in the path is only a routing hint.\n let call = m?.call?.id ? await db.getCall(String(m.call.id)) : null;\n if (!call) call = await db.latestCall(routeTenant);\n const secret = c.req.header(\"x-vapi-secret\") ?? \"\";\n if (!call || !call.webhook_secret || secret !== call.webhook_secret) {\n console.log(\"webhook: bad/missing secret for\", routeTenant, m?.call?.id ?? \"(no call id)\");\n return c.json({ error: \"bad secret\" }, 401);\n }\n const tenant = call.tenant;\n const callId = call.id;\n console.log(\"webhook:\", tenant, m.type, m.status ?? m.transcriptType ?? \"\", String(m.transcript ?? \"\").slice(0, 120));\n\n const mtype = String(m.type ?? \"\");\n if (mtype.startsWith(\"transcript\")) {\n // type is \"transcript\" or the documented variant 'transcript[transcriptType=\"final\"]'\n const ttype = m.transcriptType ?? (mtype.includes(\"final\") ? \"final\" : \"partial\");\n if (ttype === \"final\") {\n await db.addEvent(tenant, callId, \"transcript\", m.role ?? \"user\", m.transcript ?? \"\");\n await db.setPartial(callId, null);\n } else {\n await db.setPartial(callId, (m.role ?? \"?\") + \": \" + (m.transcript ?? \"\"));\n }\n } else if (mtype === \"conversation-update\") {\n // Fallback transcript source: backfill any user lines the transcript events missed.\n const conv = (m.conversation ?? m.messages ?? []) as Array>;\n const users = conv\n .filter((x) => x.role === \"user\")\n .map((x) => String(x.message ?? x.content ?? \"\"))\n .filter(Boolean);\n if (users.length) {\n const seen = new Set(\n (await db.recentEvents(callId, 80))\n .filter((e) => e.type === \"transcript\" && e.role === \"user\")\n .map((e) => e.content),\n );\n for (const u of users) {\n if (!seen.has(u)) {\n await db.addEvent(tenant, callId, \"transcript\", \"user\", u);\n seen.add(u);\n }\n }\n }\n } else if (mtype === \"status-update\") {\n await db.setCallStatus(callId, m.status ?? \"unknown\", m.endedReason ?? null);\n await db.addEvent(tenant, callId, \"status\", \"system\", String(m.status) + (m.endedReason ? \" (\" + m.endedReason + \")\" : \"\"));\n } else if (mtype === \"end-of-call-report\") {\n // Analysis (summary + structuredData) is computed at end-of-call and rides\n // along in this report. await_report waits for the \"analysis\" event below.\n await db.setCallStatus(callId, \"ended\", m.endedReason ?? null);\n const analysis = m?.analysis ?? {};\n const summary = analysis.summary ?? m.summary ?? \"call ended\";\n await db.addEvent(tenant, callId, \"report\", \"system\", String(summary).slice(0, 1000));\n if (analysis.structuredData !== undefined && analysis.structuredData !== null) {\n await db.addEvent(tenant, callId, \"analysis\", \"system\", JSON.stringify(analysis.structuredData));\n }\n }\n return c.json({});\n});\n\n// ---------- landing + live feed ----------\nfunction landing(origin: string): string {\n return `call-my-human — send an agent to phone your human (public remote MCP server)\n\nDispatch an autonomous voice agent to call your human, run the whole\nconversation itself, and hand you back a structured report. You brief it once\n(objective + context + questions); it talks, hangs up, and you collect the\nresult. The agent has a real LLM brain running in-stack, so the call is\nlow-latency and natural — you don't drive it turn by turn.\n\nTools: dispatch_call · await_report · listen · say · get_status · end_call.\nThe headline loop is just dispatch_call → await_report. listen/say are optional\nsupervisor hooks (eavesdrop on, or barge into, a live call).\n\nBring your own keys from https://dashboard.vapi.ai (private key + phone number ID).\nThe agent's brain uses your Vapi account's bundled OpenAI model (gpt-4o by\ndefault; override with &model=). No extra LLM key needed for OpenAI models.\n\nConnect from Claude Code:\n\n claude mcp add --transport http call-my-human \\\\\n \"${origin}/mcp?vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B15551234567\"\n\nNo MCP client? The tools speak plain REST with the same query params:\n POST ${origin}/api/call/dispatch {\"objective\": \"...\", \"context\": \"...\", \"questions\": [\"...\"]}\n POST ${origin}/api/call/report {\"wait_seconds\": 25} (long-polls; use a 35s timeout, repeat until done:true)\n POST ${origin}/api/call/listen {\"wait_seconds\": 25} (optional: eavesdrop live)\n POST ${origin}/api/call/say {\"text\": \"...\", \"end_call_after\": false} (optional: barge in)\n POST ${origin}/api/call/status · POST ${origin}/api/call/end\n\nNotes:\n • human_number is the default number dispatch_call dials — your cell. The +\n MUST be URL-encoded as %2B.\n • Works in any MCP client that takes a URL (claude.ai custom connectors too).\n • Watch a call live: dispatch_call returns a watch_url.\n • Free Vapi numbers are flaky (occasional dead bridges, a DAILY outbound\n limit). Import a Twilio number into Vapi for reliability.\n • Privacy: your key is read per-request and never stored. Call transcripts\n and reports ARE stored in this val's SQLite, scoped to a hash of your key.\n Want full custody? Remix this val — it has no other config.\n\nSibling: dcm31/call-my-human-no-middle-agent is the puppet version (you drive every word).\nSource + README: https://www.val.town/x/dcm31/call-my-human\n`;\n}\n\nconst PAGE = `\n\n\ncall-my-human — live\n\n\n\n

📞 call-my-human no call

\n\n`;\n\n// Favicon: the 📞 emoji as an SVG. Browsers (and connector clients that read the\n// origin's favicon to show the server's icon) fetch this.\nconst FAVICON =\n ``;\nconst favicon = () =>\n new Response(FAVICON, {\n headers: { \"content-type\": \"image/svg+xml; charset=utf-8\", \"cache-control\": \"public, max-age=86400\" },\n });\napp.get(\"/favicon.ico\", favicon);\napp.get(\"/favicon.svg\", favicon);\n\n// ---------- agent-setup instructions (file-backed) ----------\n// The splash's \"Copy agent instructions\" button copies this. It lives in an\n// editable markdown file (skills/call-my-human-setup/SKILL.md) so the copy and the\n// docs stay in one place. Read once at first use; a read failure falls back so the\n// page — and the MCP endpoint — never break.\nconst SETUP_FALLBACK =\n \"# Set up Call My Human 📞\\n\\nAdd this as a custom MCP connector (BYO free Vapi key from https://dashboard.vapi.ai):\\n\\n https://call-my-human.val.run/mcp?vapi_key=YOUR_KEY&phone_number_id=YOUR_PHONE_ID&human_number=%2B1XXXXXXXXXX\\n\\nThen: dispatch_call({ objective, questions }) -> await_report. Full docs + source: https://www.val.town/x/dcm31/call-my-human\";\nlet _setupMd: string | null = null;\nasync function setupInstructions(): Promise {\n if (_setupMd != null) return _setupMd;\n try {\n _setupMd = await readFile(\"/skills/call-my-human-setup/SKILL.md\", import.meta.url);\n } catch (e) {\n console.log(\"setup-md read failed, using fallback:\", String(e).slice(0, 140));\n _setupMd = SETUP_FALLBACK;\n }\n return _setupMd;\n}\n\nconst escapeHtml = (s: string) => s.replace(/&/g, \"&\").replace(//g, \">\");\n\n// Polished splash for humans (browsers). sleek-ui: one bold move (the aqua),\n// everything else quiet — system font, stark ink, a single primary CTA.\nfunction splashPage(md: string): string {\n return `\n\n\nCall My Human\n\n\n\n\n\n\n

📞

Remote MCP Connector

Call My Human

\n \n View source\n

Agents: /agent-instructions

${escapeHtml(md)}

\n`;\n}\n\napp.get(\"/source\", (c) => c.redirect(parseVal().links.self.val));\n\n// Agent-readable setup instructions — the SAME SKILL.md the splash copies. An\n// agent handed the bare URL, or these explicit paths, gets the exact paste block\n// as plain text. (/llms.txt is the emerging convention agents probe for.)\nconst serveAgentInstructions = async (c: any) => c.text(await setupInstructions());\napp.get(\"/agent-instructions\", serveAgentInstructions);\napp.get(\"/llms.txt\", serveAgentInstructions);\n\n// Full technical landing (MCP + REST endpoints) for anyone who wants it.\napp.get(\"/help\", (c) => c.text(landing(new URL(c.req.url).origin)));\n\napp.get(\"/\", async (c) => {\n if (c.req.query(\"c\")) return c.html(PAGE);\n // Browsers get the splash; everything else (agents, curl) gets the setup\n // instructions as plain text, so the bare URL \"just works\" when given to an agent.\n if ((c.req.header(\"accept\") ?? \"\").includes(\"text/html\")) {\n return c.html(splashPage(await setupInstructions()));\n }\n return serveAgentInstructions(c);\n});\n\nexport default app.fetch;\n",[167],"818a99fd-3dc5-4b41-b069-82b01379dbaf","db.ts","019ec890-2328-734f-a4be-e46d21840215",72,["D",1782226714613],"// SQLite layer: calls + an append-only event feed per call, scoped by tenant.\n// A tenant is a hash of the caller's Vapi key (full SHA-256) — every connecting\n// user gets their own call history on this shared public server. The tenant is\n// internal-only: it never appears in a shareable URL. Per-call random secrets\n// (watch_token, webhook_secret) gate the watch page and inbound webhooks instead.\n\nimport { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\n\nawait sqlite.batch([\n `CREATE TABLE IF NOT EXISTS calls (\n id TEXT PRIMARY KEY,\n tenant TEXT,\n control_url TEXT,\n status TEXT DEFAULT 'starting',\n ended_reason TEXT,\n live_partial TEXT,\n cursor INTEGER DEFAULT 0,\n number TEXT,\n watch_token TEXT,\n webhook_secret TEXT,\n created_at TEXT\n )`,\n `CREATE TABLE IF NOT EXISTS events (\n id INTEGER PRIMARY KEY AUTOINCREMENT,\n tenant TEXT,\n call_id TEXT,\n type TEXT,\n role TEXT,\n content TEXT,\n created_at TEXT\n )`,\n // Anonymous, append-only tally: one row per dispatched call, JUST a timestamp.\n // No tenant, no number, no content — it's a pure volume counter (calls/day).\n // The retention purge never touches this table, so the daily history survives\n // even after the transcripts/events for a call are deleted.\n `CREATE TABLE IF NOT EXISTS call_log (\n id INTEGER PRIMARY KEY AUTOINCREMENT,\n at TEXT\n )`,\n]);\n// Older/remixed schemas may predate some columns — add them if missing.\nfor (const sql of [\n `ALTER TABLE calls ADD COLUMN tenant TEXT`,\n `ALTER TABLE calls ADD COLUMN watch_token TEXT`,\n `ALTER TABLE calls ADD COLUMN webhook_secret TEXT`,\n `ALTER TABLE events ADD COLUMN tenant TEXT`,\n]) {\n try {\n await sqlite.execute(sql);\n } catch { /* already there */ }\n}\n\nconst now = () => new Date().toISOString();\n\n// Statuses that mean a call is still live. Shared with tools.ts so the retention\n// purge and the dispatch concurrency guard agree on what \"active\" means.\nexport const ACTIVE_STATUSES = [\"starting\", \"queued\", \"ringing\", \"in-progress\", \"forwarding\"];\n\n// A 128-bit random hex token — used as a per-call watch capability and webhook secret.\nexport function randomToken(bytes = 16): string {\n const a = new Uint8Array(bytes);\n crypto.getRandomValues(a);\n return [...a].map((b) => b.toString(16).padStart(2, \"0\")).join(\"\");\n}\n\nexport interface CallRow {\n id: string;\n tenant: string;\n control_url: string;\n status: string;\n ended_reason: string | null;\n live_partial: string | null;\n cursor: number;\n number: string;\n watch_token: string | null;\n webhook_secret: string | null;\n created_at: string;\n}\n\nexport interface EventRow {\n id: number;\n tenant: string;\n call_id: string;\n type: string;\n role: string;\n content: string;\n created_at: string;\n}\n\nexport async function insertCall(\n tenant: string,\n id: string,\n controlUrl: string,\n number: string,\n watchToken: string,\n webhookSecret: string,\n) {\n await sqlite.execute({\n sql: `INSERT INTO calls (id, tenant, control_url, status, cursor, number, watch_token, webhook_secret, created_at)\n VALUES (?, ?, ?, 'starting', 0, ?, ?, ?, ?)`,\n args: [id, tenant, controlUrl, number, watchToken, webhookSecret, now()],\n });\n}\n\nexport async function latestCall(tenant: string): Promise {\n const r = await sqlite.execute({\n sql: `SELECT * FROM calls WHERE tenant = ? ORDER BY created_at DESC LIMIT 1`,\n args: [tenant],\n });\n return (r.rows[0] as unknown as CallRow) ?? null;\n}\n\nexport async function getCall(id: string): Promise {\n const r = await sqlite.execute({ sql: `SELECT * FROM calls WHERE id = ?`, args: [id] });\n return (r.rows[0] as unknown as CallRow) ?? null;\n}\n\nexport async function setCallStatus(id: string, status: string, endedReason?: string | null) {\n await sqlite.execute({\n sql: `UPDATE calls SET status = ?, ended_reason = COALESCE(?, ended_reason) WHERE id = ?`,\n args: [status, endedReason ?? null, id],\n });\n}\n\nexport async function setPartial(id: string, partial: string | null) {\n await sqlite.execute({ sql: `UPDATE calls SET live_partial = ? WHERE id = ?`, args: [partial, id] });\n}\n\nexport async function setCursor(id: string, cursor: number) {\n await sqlite.execute({ sql: `UPDATE calls SET cursor = ? WHERE id = ?`, args: [cursor, id] });\n}\n\nexport async function addEvent(tenant: string, callId: string, type: string, role: string, content: string) {\n await sqlite.execute({\n sql: `INSERT INTO events (tenant, call_id, type, role, content, created_at) VALUES (?, ?, ?, ?, ?, ?)`,\n args: [tenant, callId, type, role, content, now()],\n });\n}\n\nexport async function eventsAfter(callId: string, afterId: number): Promise {\n const r = await sqlite.execute({\n sql: `SELECT * FROM events WHERE call_id = ? AND id > ? ORDER BY id ASC`,\n args: [callId, afterId],\n });\n return r.rows as unknown as EventRow[];\n}\n\nexport async function hasTranscript(callId: string): Promise {\n const r = await sqlite.execute({\n sql: `SELECT 1 FROM events WHERE call_id = ? AND type = 'transcript' LIMIT 1`,\n args: [callId],\n });\n return r.rows.length > 0;\n}\n\nexport async function latestEventOfType(callId: string, type: string): Promise {\n const r = await sqlite.execute({\n sql: `SELECT * FROM events WHERE call_id = ? AND type = ? ORDER BY id DESC LIMIT 1`,\n args: [callId, type],\n });\n return (r.rows[0] as unknown as EventRow) ?? null;\n}\n\nexport async function transcriptLines(callId: string, limit = 200): Promise {\n const r = await sqlite.execute({\n sql: `SELECT * FROM events WHERE call_id = ? AND type = 'transcript' ORDER BY id ASC LIMIT ?`,\n args: [callId, limit],\n });\n return r.rows as unknown as EventRow[];\n}\n\nexport async function recentEvents(callId: string, limit = 100): Promise {\n const r = await sqlite.execute({\n sql: `SELECT * FROM events WHERE call_id = ? ORDER BY id DESC LIMIT ?`,\n args: [callId, limit],\n });\n return (r.rows as unknown as EventRow[]).reverse();\n}\n\n// ---------- anonymous volume tally ----------\n\n// Record that a call happened. Just a timestamp — no tenant, number, or content.\nexport async function logCall(): Promise {\n await sqlite.execute({ sql: `INSERT INTO call_log (at) VALUES (?)`, args: [now()] });\n}\n\n// Calls per UTC day, most recent first. Aggregate-only; no per-user breakdown.\nexport async function callsPerDay(days = 90): Promise> {\n const r = await sqlite.execute({\n sql: `SELECT substr(at, 1, 10) AS day, COUNT(*) AS count\n FROM call_log GROUP BY day ORDER BY day DESC LIMIT ?`,\n args: [days],\n });\n return r.rows as unknown as Array<{ day: string; count: number }>;\n}\n\n// ---------- retention purge ----------\n\n// Delete finished calls (and their events) older than `retentionDays`. NEVER\n// touches call_log (the daily tally), and NEVER an in-flight call: the age cutoff\n// alone is ~1000x the 15-min max call length, and we also exclude ACTIVE_STATUSES\n// as belt-and-suspenders. Returns how many call rows were removed.\nexport async function purgeExpired(retentionDays: number): Promise<{ calls: number }> {\n const cutoff = new Date(Date.now() - retentionDays * 86400_000).toISOString();\n const placeholders = ACTIVE_STATUSES.map(() => \"?\").join(\", \");\n const where = `created_at < ? AND status NOT IN (${placeholders})`;\n const r = await sqlite.execute({\n sql: `SELECT COUNT(*) AS n FROM calls WHERE ${where}`,\n args: [cutoff, ...ACTIVE_STATUSES],\n });\n const n = Number((r.rows[0] as any)?.n ?? 0);\n if (n > 0) {\n await sqlite.batch([\n {\n sql: `DELETE FROM events WHERE call_id IN (SELECT id FROM calls WHERE ${where})`,\n args: [cutoff, ...ACTIVE_STATUSES],\n },\n { sql: `DELETE FROM calls WHERE ${where}`, args: [cutoff, ...ACTIVE_STATUSES] },\n ]);\n }\n return { calls: n };\n}\n",[175],"73eb705b-021c-4668-9a35-3d0e7c625c6a","cleanup.ts","0f18da28-6f14-11f1-8760-1607ee4eb77e",75,["D",1782226736836],0,"interval","// Daily retention sweep. Deletes finished calls + their transcripts/events older\n// than RETENTION_DAYS so this shared public val's SQLite doesn't grow forever.\n//\n// Safety: the age cutoff is days, while the longest a call can run is 15 minutes\n// (max_minutes is capped in tools.ts), so nothing in flight can ever be touched —\n// and purgeExpired also explicitly excludes ACTIVE statuses. The anonymous\n// calls/day tally (call_log) is NEVER purged, so volume history is permanent.\n//\n// Set RETENTION_DAYS to override the 14-day default (per-instance; a remix can\n// keep more or less). Runs once a day — see this file's interval settings.\n\nimport * as db from \"./db.ts\";\n\nconst RETENTION_DAYS = Math.max(Number(Deno.env.get(\"RETENTION_DAYS\") ?? 14), 1);\n\nexport default async function () {\n const { calls } = await db.purgeExpired(RETENTION_DAYS);\n console.log(`cleanup: purged ${calls} call(s) older than ${RETENTION_DAYS}d (call_log tally untouched)`);\n}\n",[182],"a0f25665-e6f8-4ffe-a0e6-90d9ffa11362",["D",1782226793979],[28],"574d4a15-ac47-4b66-ace0-427151103d23","skills","f9f9dc1b-d70d-411d-be18-75bcb316622f",60,["D",1782166791021],"directory",[194],"pulls","branches","forks","envVars",{"_207":209,"_210":186},"connections",{"_207":186,"_208":186},"count","unfilledCount",4,"emptyCount",{"_42":193,"_32":194,"_30":195,"_141":-5,"_40":41,"_39":196,"_143":59,"_37":258,"_13":186,"_33":198,"_34":-5,"_24":25,"_148":-5,"_222":259,"_27":260},{"_42":254,"_32":253,"_30":246,"_141":195,"_40":41,"_39":196,"_143":59,"_37":255,"_13":186,"_33":198,"_34":-5,"_24":25,"_148":-5,"_222":256,"_27":257},{"_42":243,"_32":244,"_30":245,"_141":246,"_40":41,"_39":247,"_143":25,"_37":248,"_13":249,"_33":20,"_34":250,"_24":25,"_148":-5,"_222":251,"_27":252},{"_42":190,"_32":28,"_30":31,"_141":-5,"_40":41,"_39":14,"_143":25,"_37":240,"_13":155,"_33":20,"_34":36,"_24":25,"_148":-5,"_222":241,"_27":242},{"_42":181,"_32":182,"_30":183,"_141":-5,"_40":41,"_39":184,"_143":25,"_37":237,"_13":186,"_33":187,"_34":188,"_24":25,"_148":-5,"_222":238,"_27":239},{"_42":174,"_32":175,"_30":176,"_141":-5,"_40":41,"_39":177,"_143":25,"_37":234,"_13":163,"_33":146,"_34":179,"_24":25,"_148":-5,"_222":235,"_27":236},{"_42":166,"_32":167,"_30":57,"_141":-5,"_40":41,"_39":168,"_143":25,"_37":231,"_13":170,"_33":171,"_34":172,"_24":25,"_148":44,"_222":232,"_27":233},{"_42":158,"_32":159,"_30":160,"_141":-5,"_40":41,"_39":161,"_143":59,"_37":228,"_13":163,"_33":146,"_34":164,"_24":25,"_148":-5,"_222":229,"_27":230},{"_42":150,"_32":151,"_30":152,"_141":-5,"_40":41,"_39":153,"_143":25,"_37":225,"_13":155,"_33":146,"_34":156,"_24":25,"_148":-5,"_222":226,"_27":227},{"_42":138,"_32":139,"_30":140,"_141":-5,"_40":41,"_39":142,"_143":59,"_37":221,"_13":145,"_33":146,"_34":147,"_24":25,"_148":-5,"_222":223,"_27":224},["D",1782163327615],"children",[],[139],["D",1782226727210],[],[151],["D",1782155923638],[],[159],["D",1782226751300],[],[167],["D",1782226714613],[],[175],["D",1782226736836],[],[182],["D",1782226793979],[],[28],"6c9b177c-8b73-4d3c-8a4b-a0b6aaa8d7a1","SKILL.md","7ca27d9c-6e88-11f1-9be9-1607ee4eb77e","09f7ea79-9c23-4f81-bb4e-0847987d112b",69,["D",1782222605518],3,"---\nname: call-my-human-setup\ndescription: Set up the Call My Human MCP connector so an AI agent can place a real phone call to its human — ask a question, get a decision, run a quick errand call — and get a structured report back. Use when a human pastes this and says \"set this up for me.\"\n---\n\n# Set up Call My Human 📞\n\nYou're being asked to wire up **Call My Human** — a public remote MCP server that lets you (the agent) dispatch a voice agent to phone your human, run the whole call itself, and hand back a structured report (summary + answer to each question you asked). It's BYOK: it runs on your human's own Vapi account, and the server stores only transient call state.\n\nWalk your human through this. They don't need anything beforehand.\n\n## 1. Get free Vapi credentials (~2 min)\n\nSend them to **https://dashboard.vapi.ai**, have them sign up (free), then grab two things:\n\n- their **private API key**\n- a **phone number ID** — Vapi hands out a free US number. The free tier covers **US calling, ~10 outbound calls/day** — plenty to try this out.\n\nNo separate AI key is needed; the call's brain runs on Vapi's bundled model.\n\n## 2. Connect\n\n**If you're a CLI agent (Claude Code, Cursor, Codex)** — you can add it yourself:\n\n```\nclaude mcp add --transport http call-my-human \\\n \"https://call-my-human.val.run/mcp?vapi_key=THEIR_KEY&phone_number_id=THEIR_PHONE_ID&human_number=%2B1XXXXXXXXXX\"\n```\n\n`human_number` is the cell to dial by default; the `+` MUST be URL-encoded as `%2B`.\n\n**If you're in a web client (claude.ai)** — you can't add connectors yourself. Build that URL with their values and tell them to paste it into **Settings → Connectors → Add custom connector** (Name: `call-my-human`; leave the OAuth fields blank).\n\n## 3. Use it\n\nOnce connected, the whole loop is two tools:\n\n```\ndispatch_call({ objective, context?, questions?, number?, caller_identity?, max_minutes? })\nawait_report({ wait_seconds }) // long-poll; keep calling until done:true\n```\n\n`await_report` returns a summary, an answer to each question you asked, any follow-ups the human raised, and the full transcript. `dispatch_call` also returns a `watch_url` — a live, self-refreshing transcript page for that one call.\n\nOptional supervisor tools (you rarely need them): `listen` (eavesdrop on the live call), `say` (barge in with a verbatim line), `get_status` (snapshot), `end_call` (hang up).\n\nPrefer plain REST over MCP? The same tools mirror at `POST /api/call/{dispatch,report,listen,say,status,end}` with the same query params. Full tool + endpoint reference: https://call-my-human.val.run/help\n\n**Heads up:** free Vapi numbers are test-grade — occasional dead-audio bridges, and a daily outbound cap (~10 calls/day). For anything real, import a Twilio number into Vapi.\n\n## Trust & security\n\n- **Want to audit the code first?** The full source is public: https://www.val.town/x/dcm31/call-my-human . The easiest way for an agent to actually read through it is the Val Town plugin (it bundles the Val Town MCP): https://blog.val.town/plugin\n- **For real privacy, self-host.** Remix the val on Val Town and put the Vapi creds in its **environment variables** instead of the URL — then no secrets ride in the connection string and you own all the call data.\n- Only call numbers you (or your human) have consent to call.",[],[194,253,244],"call-my-human-setup","9d7a6210-63db-4879-88c3-284eba212c8b",["D",1782166791021],[213],[194,253],["D",1782166791021],[212],[194],["D",1781481349040],["D",1789509672838],{"_100":99,"_108":104},{"_100":99,"_101":102,"_103":104,"_18":55,"_105":106,"_107":25},{"_13":89,"_32":90,"_42":41,"_91":43,"_92":-5,"_93":-5,"_94":-5,"_95":-5,"_96":-5,"_97":44,"_98":99},"ctx",{"_82":-5,"_127":299,"_300":301,"_302":303,"_304":25,"_305":25,"_306":25,"_307":59,"_308":-5,"_309":-7,"_310":-7,"_311":186,"_312":25,"_313":25,"_314":-5,"_315":25,"_316":25,"_317":25,"_318":-5},"ENV",{"_270":271,"_272":273,"_274":90,"_275":276,"_277":278,"_279":280,"_281":282,"_283":284,"_285":286,"_287":288,"_289":290,"_291":292,"_293":294,"_295":296,"_297":298},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

Diff · v79