Lifescored MCP Server

life. scored.

You are already a number.

A credit score, an actuarial row, a callback probability — this app rebuilds those numbers in the open: every rule cited or flagged as a guess, every weight visible and editable. Transparency, not judgment.

Stack

• Svelte 5 + SvelteKit, prerendered static pages, deployed to Cloudflare Workers
• One dynamic endpoint: POST /api/narrative — KV-cached Gemini narratives with a deterministic local fallback (the app is fully functional with no API key at all)
• All user data stays in the browser (localStorage + URL-fragment share links)
• Traffic is measured with Cloudflare Web Analytics — cookieless and aggregate; it cannot see your inputs, which never leave your device anyway.

Develop

npm install
npm run dev      # app at localhost:5173 (narrative falls back to local composer)
npm test         # vitest suite: rulebook invariants, engine, codec, worker handler
npm run check    # svelte-check

Deploy (Cloudflare free tier)

Deploys run automatically from GitHub Actions on every push to main (.github/workflows/deploy.yml): the workflow runs the full test suite + typecheck, and only then builds and deploys.

One-time setup:

1. Create the KV namespace and paste its id into wrangler.jsonc:

   npx wrangler kv namespace create NARRATIVE_KV
   

2. Set the Gemini key as a Worker secret (persists across deploys; optional — omit to run AI-free with the local narrative fallback):

   npx wrangler secret put GEMINI_API_KEY
   

3. Add two repository secrets in GitHub (Settings → Secrets and variables → Actions):

   • CLOUDFLARE_API_TOKEN — a token with the Edit Cloudflare Workers template
   • CLOUDFLARE_ACCOUNT_ID — your account id

Push to main and the action ships it. To deploy by hand instead:

   npm run deploy

Deploy your own

Fork the repo, do the one-time setup above with your own Cloudflare account, and push. The app is fully functional with no Gemini key at all.

Editing the rulebook

Rules live in src/lib/rulebook/<domain>.ts. Each rule is declarative: logic, evidence tag (SOURCED/SPECULATIVE), citation with access date, a pure position(inputs) (the measured fact, normalized to a numeric position), declared bounds (negative floors only where the cited system itself subtracts; the wealth rules are uncapped above), and a weightRationale justifying its weight against the 1.0× income baseline; the engine computes points = round(position × weight). Add a rule, and the UI, weight editor, share codec, and about-page source list pick it up automatically. npm test enforces the invariants (bounds, integer scores, citation present).

Use it from an agent (MCP or skill)

The whole rulebook + exact math is published so any agent can compute a score on its own side — nothing about a person is ever sent to lifescored.com.

• MCP server — com.lifescored/mcp on the official registry; endpoint https://lifescored.com/mcp (stateless Streamable HTTP, no auth, accepts no personal data). Tools: get_rulebook, get_input_schema, get_methodology, how_to_give_feedback.
• Skill — a drop-in skills/lifescored/ skill. Copy it into your agent's skills directory (e.g. ~/.claude/skills/). It prefers the MCP and falls back to rules.json.
• Raw data — https://lifescored.com/rules.json and /llms.txt.

Contributing

Every commit must reference a GitHub issue number, e.g. feat: add presets (#2) or refs #5. A commit-msg hook enforces this automatically — it is installed via core.hooksPath pointing to .githooks/, which activates on npm install (the prepare script runs git config core.hooksPath .githooks). Merge commits and reverts are exempt. To find or open an issue: gh issue list / gh issue create.

License

MIT — see LICENSE.