Pulse

A public repair journal for people and browser agents. Record a problem, proposed checks, human observations, attempted fixes, and the final outcome. Search the saved history later. Every step works through ordinary forms; compatible agents can use the same database through ten WebMCP tools.

Open Pulse · Browse examples · Start a repair · Agent guide

Use it

1. Search community cases by product, symptom, outcome, or difficulty. Switch Record source to Examples / practice to explore fictional records.
2. Choose Start a Repair. Describe the object and the observed problem, then select its safety classification. Select practice case when trying fictional data.
3. Open the case. Use Continue this case to propose a check, record an observation, and document an attempt and its parts/cost.
4. Record the outcome, total cost in USD, whole minutes spent, and notes. Failures and professional-service decisions are useful outcomes too. Saving an outcome replaces the previous outcome; checks and attempts remain.
5. Reload to verify the saved evidence. Use Export case JSON for a readable backup.

Everything submitted is public. Avoid personal information and serial numbers. The browser that creates a case receives an opaque, HttpOnly edit cookie; the server stores only its hash. Keep the same browser profile and cookies to continue editing. Clearing cookies or moving to another device removes editing access. There is no account, recovery flow, edit-key export, or public delete action. Existing cases created before browser ownership was introduced remain readable without an edit-claim mechanism.

Reading, searching, and exporting are public. Other visitors can leave self-reported feedback, counted once per feedback type per browser. This is not a verified identity or reputation system.

Honest evidence

• The database starts with 30 clearly labeled fictional examples. New practice cases are labeled and separated with them. One precisely identified legacy automated test record is also relabeled as practice, preserving its history.
• Search defaults to community cases. Examples require the source filter; tool results include their source.
• Dashboard and homepage community totals exclude examples and practice. Outcomes are contributor reports, not independently verified physical repairs.
• Success rate is fixed/improved cases divided by all cases with an outcome. Open cases are excluded from that denominator. No outcomes means no rate.
• Median cost uses recorded outcome costs. Each case shows its actual recorded cost/time, with no invented price range.
• Database failures show an error instead of substituting fictional evidence. Statistics cover all saved cases. Search returns up to 50 matches; common-failure summaries explicitly describe their sample of up to 50 matching community cases.

Pulse organizes repair records; it does not diagnose an object or generate repair advice by itself. People supply physical observations. Cases classified professional recommended accept history and reported outcomes, but the server refuses new procedural diagnostic steps. Classification is contributor-supplied and is not a safety certification.

WebMCP

The page registers tools through document.modelContext.registerTool, with a compatibility fallback for navigator.modelContext. Browsers without the API show a clear status and retain the full form workflow. No provider key or paid model is required by Pulse; an agent is optional and supplied by the visitor's browser.

Tool	What it does
search_repairs	Search community cases; optionally request examples or all records
get_repair_case	Read the complete history, outcome notes, source, and this browser's edit permission
create_repair_case	Publish a case; practice: true marks fictional data
add_diagnostic_step	Propose a check on a case this browser owns; blocked for professional-risk cases
add_diagnostic_result	Save a person's reported observation against a step
record_repair_attempt	Record an attempted fix, parts, cost, and difficulty
record_repair_outcome	Save or correct the current reported outcome
mark_case_helpful	Save helpful/worked-for-me/did-not-work feedback once per type and browser
list_common_failures	Summarize up to 50 matching community cases, with sample size disclosed
get_repair_statistics	Read community totals and the separate example/practice count

Tool responses are complete JSON, not strings cut at an arbitrary character limit. Community text is untrusted data. Tool metadata is static; input schemas and server validation enforce field limits. UI and tool writes share the same HTTP handlers, ownership checks, and D1 database. Agents can record only observations supplied by a person; use practice cases for fictional exercises.

Run locally

Use Node.js 24 and pnpm 11.19.0 (the versions used in CI).

git clone https://github.com/agammann/pulse-webmcp.git
cd pulse-webmcp
pnpm install --frozen-lockfile
pnpm dev

Open the URL printed by the dev server. The Cloudflare plugin provides local D1 storage under .wrangler/; no API key is needed. Empty databases initialize the schema and example corpus automatically.

To run the built Worker:

pnpm build
pnpm start --port 3015

The public site uses the same build with a durable D1 binding named DB, declared in .openai/hosting.json. Local D1 is separate from the public database. Sites deploys dist/ from a source commit; GitHub Actions verifies source changes but does not deploy them automatically. Hosting a separate copy requires your own compatible Cloudflare/Sites environment and may have hosting costs.

Verify changes

pnpm test
pnpm lint
pnpm typecheck
pnpm exec playwright install chromium
pnpm build
pnpm test:e2e

Unit tests cover ranking, boundaries, metadata, cookies, and statistics including more than 50 cases. Browser tests run the built Worker with local D1 and cover the full ordinary-browser repair journal, reload/export, ownership, feedback deduplication, professional-case boundaries, source filtering, unavailable data, mobile pages, and all ten tool adapters. A test-only registry captures adapters in Playwright; this is separate from native WebMCP verification in a compatible browser.

Source map

• app/api/ — shared HTTP reads and writes
• lib/database.ts — D1 state and additive schema initialization
• lib/session.ts, lib/access.ts — browser edit key and write boundary
• lib/validation.ts, lib/tool-input.ts — server and tool input checks
• lib/search.ts, lib/statistics.ts — deterministic ranking and reported totals
• components/repair-editor.tsx — manual journal forms
• components/webmcp-provider.tsx, lib/webmcp-contracts.ts — native registration and ten contracts
• drizzle/0000_pulse.sql, drizzle/0001_browser_editors.sql, drizzle/0002_legacy_practice.sql — initial and additive schemas; runtime uses idempotent creation

Limits

Pulse has no contributor accounts, moderation dashboard, cross-device editing, or deletion UI. Rate limiting is basic and instance-local. Browser cookies limit accidental duplicate feedback; clearing them can create a new voter identity. Outcome corrections replace the current outcome rather than preserving revisions. Search is a deterministic in-memory ranking over loaded case rows, suited to a small collection. Public case text and safety classifications are self-reported.

MIT licensed. See LICENSE and security and safety policy.