An email-based personal AI agent built on Val Town. Interact entirely via email — send messages, tasks, ideas, and memories to memory-do@valtown.email and the agent classifies, stores, executes, and replies.

Send an email. The agent reads it, figures out what you mean, and acts on it.

You → "Remind me Thursday to call the dentist"
     → "I like chickens"
     → "What's the latest news on WebAssembly?"
     → "Every Monday, give me a summary of my active tasks"

Under the hood:

1. Classify — Claude analyzes your email and extracts structured entries (memories, tasks, ideas, people, meetings, projects)
2. Store — Everything goes into a unified SQLite database with full-text search and a relationship graph linking entries together
3. Execute — Instant tasks run immediately through an agentic loop; future/recurring tasks get scheduled
4. Reply — You get an email back with results, follow-up questions, and any generated file attachments

Emails sent via voice transcription work great. The classifier is tuned for misspellings, run-on sentences, and implicit intent — it interprets what you mean, not what you literally typed.

The agent doesn't just make a single LLM call and reply. It runs an iterative tool-use loop — Claude can call tools mid-task to gather information, search the web, query your stored knowledge, and create follow-up work.

┌─────────────────────────────────────────────┐
│  1. Load user context (memories, tasks, etc) │
│  2. Send task + context + tools to Claude     │
│  3. Claude responds:                          │
│     ├─ Final answer? → Done, send reply       │
│     └─ Tool call? → Execute tool, loop to 2   │
│  4. Safety: max 15 iterations                 │
│  5. Extract attachments, generate subject     │
└─────────────────────────────────────────────┘

Tool	What It Does
search_memories	Full-text search across all your stored entries
get_entries	Browse entries by type (all tasks, all people, all ideas, etc.)
get_linked_entries	Traverse the relationship graph between entries
store_memory	Save new facts discovered during task execution
create_task	Create follow-up tasks — instant, future, or recurring
fetch_url	Fetch live data from the web (news, APIs, weather, docs)

This means the agent can do multi-step research: search your memories, fetch a web page, cross-reference with your stored projects, save what it learned, and compose a rich reply — all in one email exchange.

Everything is an entry with a type. The system is extensible — new types need zero code changes.

Type	Example
memory	"I work at Acme Corp", "I like chickens"
task	"Remind me Thursday to call the dentist"
idea	"I have an idea for a recipe app..."
person	"My friend Sarah works at Google"
meeting	"Lunch with Bob on Friday at noon"
project	"I'm building a Chrome extension"
interaction	Every email sent/received (automatic)

Entries are linked to each other in a relationship graph — "this idea became this project", "these entries came from the same email", "this person is connected to this meeting".

Type	Behavior	Example
Instant	Executes immediately, reply in same email	"What's the weather in London?"
Future	Scheduled for a specific time	"Remind me Thursday at 10am to call dentist"
Recurring	Repeats on a schedule	"Every Monday, summarize my active tasks"

Future and recurring tasks are processed by a task runner that checks every 15 minutes.

Each user gets a persistent system prompt — a personal "CLAUDE.md" that shapes how the agent behaves for them.

Set it by emailing instructions:

• "Be more casual and use emojis"
• "Call me Paul"
• "Always use bullet points"
• "Don't be so wordy"

These are extracted automatically and applied to every future interaction.

Full bidirectional support:

• Inbound — Email attachments (CSV, JSON, text, code files, etc.) are stored and their content is fed to both the classifier and the agent as context
• Outbound — The agent can generate files (reports, CSV exports, data dumps) and send them back as email attachments

email-agent/
  README.md             ← You are here
  CLAUDE.md             — Internal dev guide and conventions
  main.ts               — HTTP admin/debug endpoint
  email-handler.ts      — Email ingestion (memory-do@valtown.email)
  task-runner.ts        — Scheduled task processor (runs every 15 min)
  lib/
    classifier.ts       — Claude-powered email → structured entries
    agent.ts            — Agentic task execution with tool-use loop
    scheduler.ts        — Natural language → ISO datetime
    attachments.ts      — Blob storage for inbound/outbound files
  db/
    schema.ts           — SQLite schema (entries + FTS5 + links)
    queries.ts          — CRUD, search, user context builder

• Runtime: Val Town (Deno-based serverless)
• LLM: Claude Sonnet via Anthropic SDK
• Database: Project-scoped SQLite with FTS5 full-text search
• Storage: Val Town Blob API for attachments
• Email: Val Town email sending/receiving

        ┌──────────────┐
        │  Incoming     │
        │  Email        │
        └──────┬───────┘
               │
        ┌──────▼───────┐
        │  Classifier   │  ← Claude: extract entries, instructions, follow-ups
        │  (1 LLM call) │
        └──────┬───────┘
               │
        ┌──────▼───────┐
        │  Store in DB  │  ← Entries, FTS sync, cross-links
        └──────┬───────┘
               │
       ┌───────┴────────┐
       │                 │
  ┌────▼─────┐    ┌─────▼──────┐
  │ Instant   │    │ Future /   │
  │ Tasks     │    │ Recurring  │
  │           │    │ Tasks      │
  │ ┌───────┐ │    │            │
  │ │ Agent │ │    │ Queued for │
  │ │ Loop  │ │    │ task-runner│
  │ │ ↻     │ │    └────────────┘
  │ └───┬───┘ │
  │     │     │
  └─────┼─────┘
        │
  ┌─────▼─────┐
  │  Reply     │  ← Results + attachments + follow-up questions
  │  Email     │
  └────────────┘

Variable	Required	Description
ANTHROPIC_API_KEY	Yes	Used by all Claude API calls (classifier, scheduler, agent)

The HTTP endpoint (main.ts) provides:

• GET / or GET /stats — System-wide entry counts, entries by type, link counts
• GET /user/:email — Full user profile with entries grouped by type

• No conversation threading — Each email is independent (though recent history is loaded as context)
• Plain text replies — No HTML email rendering
• No admin auth — HTTP endpoints are open
• No rate limiting — No flood protection
• No memory deduplication — Semantic similarity dedup not yet implemented
• No system prompt reset — Users can add instructions but can't easily clear them yet

Built on Val Town. Do what you want with it.