Research snapshot: 2026-09-06.

Ranking target: normal ChatGPT mobile/web, no user PC, recurring free allowance, deterministic browser control, strong isolation, minimal browser-open time and zero-surprise-cost preference.

Runtime state: ACCOUNT_KEY_REQUIRED / live browser UNVERIFIED.

Current official Free plan:

• $0/month plus usage costs
• $5/month free credits
• 5 concurrent browsers
• live view
• stealth mode including CAPTCHA solver/proxy
• managed auth (3 connections)
• one-day session replays
• uploads/downloads
• no card required to get started
• usage billed by GB-second

Technical fit:

• official Val Town browser-automation guide/starter
• @onkernel/sdk
• server-side Playwright execution in the browser VM
• on-demand browsers.create() is explicitly recommended for occasional/bursty/one-off work
• browser is destroyed in finally

Adapter: providers/kernel.ts. Required env var: KERNEL_API_KEY.

Official sources:

• https://www.kernel.sh/pricing
• https://www.kernel.sh/docs
• https://docs.val.town/guides/browser-automation/kernel

Runtime state: ACCOUNT_TOKEN_REQUIRED / live browser UNVERIFIED.

Current official Free plan:

• 1,000 units/month
• 2 concurrent browsers
• 2-minute max session
• no card required
• Chrome / WebKit / Firefox
• automatic CAPTCHA solving available
• persisted sessions/replays

Unit model:

• browser time: 1 unit per 30-second connection block
• partial blocks round up
• successful CAPTCHA solve: 10 additional units
• residential proxy: 6 units/MB
• datacenter proxy: 2 units/MB

September 2026 provider control:

• account overage spend limit can be set to $0
• this should be mandatory before promotion to fallback OPERATIONAL under this project's zero-surprise-cost policy

Adapter: providers/browserless.ts. Required env var: BROWSERLESS_TOKEN. Optional: BROWSERLESS_BASE_URL.

Official sources:

• https://www.browserless.io/pricing
• https://docs.browserless.io/overview/unit-consumption
• https://www.browserless.io/changelog/2026-09

State: researched, not implemented.

Val Town has an official Browserbase integration guide and starter. It remains behind Kernel/Browserless because the current combination above better matches recurring-free allowance + CAPTCHA/anti-bot + cost control for this project.

Source:

• https://docs.val.town/guides/browser-automation/browserbase

State: researched, not implemented.

Useful properties:

• open-source browser API
• Puppeteer/Playwright
• official Val Town integration guide

Lower priority because one-time starter credit is less aligned with recurring zero-cost steady state than Kernel/Browserless.

Source:

• https://docs.val.town/guides/browser-automation/steel

1. Explicit configured preferredProvider is honored unless its circuit is OPEN.
2. Otherwise Kernel is primary.
3. Browserless fallback is allowed only for read_only.
4. Provider BLOCKED for missing configuration does not increment health failures.
5. Provider FAIL increments consecutive failures.
6. Three consecutive FAILs open the provider circuit for 15 minutes.
7. PASS/PARTIAL resets provider failure state.
8. Never automatically retry a write on another provider.
9. Keep reasoning outside the browser; send only a short deterministic DSL plan.
10. No automatic cron runs browsers.

A provider cannot become OPERATIONAL from docs/pricing/module-load alone. Required evidence:

• env-var key exists
• credential authenticates
• real browser is created
• network preflight passes
• read-only smoke plan executes
• explicit expectations PASS
• attempt is persisted
• provider health remains CLOSED
• actual account consumption is observed if available

Until all of these are true, state remains UNVERIFIED even if the SDK call returns successfully.