Kill My Idea

Describe a startup idea. Jev decides: KILL IT, FIX IT or SHIP IT.

No generative LLM. One request asks TypeSafe Jev 10 questions in parallel: 8 plain indie-hacker questions scored 0–4, plus the category and whether the idea is understandable.

idea → Jev (8 scores) → each × 25 → weighted average → clarity gate → KILL / FIX / SHIP

1. Install

npm install

2. Environment

cp .env.example .env

TYPESAFE_API_KEY=your_key

The key is only read by api/evaluate.ts and never reaches the browser.

Successful evaluations are archived in SQLite at data/analytics.sqlite by default. Set ANALYTICS_DB_PATH to an absolute path on a persistent disk in production. Users can opt out per request with the checkbox below private history.

SITE_URL (optional) sets the absolute URL used in the canonical and OG/Twitter tags, robots.txt, sitemap.xml, /mcp.md and llms.txt (generated at build time). It defaults to https://killmyidea.biz. The image itself is public/og.png (1200×630).

For UI work without a key, set TYPESAFE_MOCK=1. The API then returns deterministic fake answers labelled MOCK DATA in the raw-data panel. Mock mode is ignored in production.

3. Run locally

npm run dev          # http://127.0.0.1:5317 (serves /api/* and /mcp/ too)
npm test             # unit tests
npm run benchmark    # balanced scoring benchmark; requires the app to be running

Every result has a collapsed "How Jev decided" panel with raw Jev answers, probabilities, confidence, the average, the verdict, latency and token usage. ?debug=1 opens it by default.

4. Build

npm run build        # typecheck + static build in dist/

5. Deploy (Laravel Forge)

The site runs on Forge as a Node server: npm run build for the static assets in dist/, then npm run build:server && npm start to serve the API endpoints. Set TYPESAFE_API_KEY in the site's environment file.

SQLite needs a persistent local disk, which the Forge server provides directly (data/analytics.sqlite); set ANALYTICS_DB_PATH to an absolute path on that disk in production.

Because the SPA is served with an nginx fallback (try_files ... /index.html), every unknown path returns the index page with HTTP 200. Bots probing random URLs therefore show up as pageviews in DataFast. To stop that, change the fallback in the site's nginx configuration (Forge → site → Edit Nginx Configuration) to return 404 for unknown paths if the app has no real routes:

location / {
    try_files $uri $uri/ =404;
}

Domain move (killmyidea.stemonte.io → killmyidea.biz)

Both domains belong to the same Forge site. The old one answers every request with a 301 to the same path and query on killmyidea.biz, via deploy/nginx-old-domain-redirect.conf, pasted into the site's nginx configuration inside the server block. Keep that redirect for at least a year.

Idea history lives in localStorage, which is per origin. The 301 adds #moved; the inline script in index.html then makes one hop to public/history-transfer.html on the old domain, which returns with its history in the URL fragment (never sent to a server) to be merged. It runs at most once per browser. Remove both once the old domain is retired.

MCP server

api/mcp.ts is a free, public MCP server at /api/mcp (Streamable HTTP, no auth). It is stateless: every POST is one JSON-RPC message answered with plain JSON, there are no sessions and GET returns 405. Tools:

• judge_idea(idea, goal?, do_not_archive?): same pipeline as /api/evaluate (api/_judge.ts), returns the verdict, score, best signal, biggest risk and the eight sub-scores, without the raw Jev response
• get_stats(): analyzed count and verdict split

MCP evaluations are archived like web ones, with source = 'mcp'.

Usage is tracked in the mcp_requests table: one row per JSON-RPC message with the method, tool name, client name and version (sent with initialize), user agent, protocol version, success or error, and latency. It never stores tool arguments or IP addresses. Summary for the last 7 days (or -- 30 for 30):

npm run mcp:usage

That report includes user agents and error messages, so it stays private. The public /stats/ page (stats/index.html + src/stats-page.tsx, linked in the footer) reads GET /api/usage (api/usage.ts, cached 5 minutes): ideas analyzed on the web and via MCP, a 30-day daily chart, tool calls, clients and verdicts. Client and tool names come from callers, so only known ones are published (clientLabel), everything else is "Other".

Docs: mcp/index.html is the static SEO page at /mcp/ (a second Vite entry); mcp/mcp.md is served as /mcp.md for agents and llms.txt is generated in vite.config.ts. Both are emitted at build time and served by the dev server.

claude mcp add --transport http killmyidea https://killmyidea.biz/api/mcp

nginx already proxies /api/ to the Node server, so no extra config is needed.

6. Scoring

src/lib/scoring.ts and src/lib/verdict.ts. That's all there is:

• each question: Jev score 0–4 × 25 → 0–100
• score = weighted average of the 8 questions (WEIGHTS)
• Make money weights Real problem and Money twice; Open source replaces Money with Adoption; Just for fun replaces Real problem and Money with Immediate appeal and Fun. The two goal-defining questions always count twice. The API takes goal (money default, open_source, fun).
• KILL below 50 · FIX 50–64 · SHIP 65+ (VERDICT_THRESHOLDS)
• if understandability is below 0.3, the raw scores are retained for analysis but the UI asks for more detail instead of presenting a score or verdict
• SCORING_VERSION identifies the exact questions, weights, thresholds and gating semantics

BEST SIGNAL / BIGGEST RISK are the highest / lowest question, with fixed copy in src/lib/copy.ts.

7. Jev questions

src/lib/questions.ts: Real problem, Clear customer, Demand, Money, Reach, Different, Buildable, Shareable, plus Adoption, Immediate appeal and Fun for goal-specific scoring. Each has five levels written as concrete situations (Jev judges each level on its own). The HTTP client is src/lib/typesafe.ts.

Layout

api/evaluate.ts           API endpoint for the website
api/mcp.ts                MCP server (JSON-RPC over POST /api/mcp)
api/_judge.ts             shared pipeline (validate → Jev → score → archive)
api/stats.ts              API endpoint (ideas submitted + verdict split, 5-min cache)
api/usage.ts              public usage numbers for /stats/ (5-min cache)
api/_analytics-db.ts      SQLite archive for evaluation requests and scores
api/_mock.ts              local-only fake answers
src/lib/questions.ts      Jev questions
src/lib/scoring.ts        0-4 → 0-100, average
src/lib/verdict.ts        KILL / FIX / SHIP thresholds
src/lib/copy.ts           fixed copy per question
src/lib/storage.ts        history (IdeaStore interface, localStorage implementation)
src/lib/share.ts          share text and X intent URL
src/lib/features.ts       public site and repository URLs
src/components/           UI
mcp/index.html            /mcp/ docs page (SEO)
mcp/mcp.md                /mcp.md docs for agents
stats/index.html          /stats/ public usage page
benchmarks/cases.json     balanced, synthetic benchmark cases
scripts/run-scoring-benchmark.mjs  benchmark runner and summary

Analytics (DataFast, cookieless)

Production builds load DataFast's cookieless script (script.cookieless.js) with the site id dfid_UvVlC3fid8kOr95YTMIho (vite.config.ts). Dev builds only track when DATAFAST_WEBSITE_ID is set; DATAFAST_WEBSITE_ID=off disables tracking. DATAFAST_DOMAIN defaults to killmyidea.biz; DATAFAST_ALLOWED_HOSTNAMES defaults to killmyidea.stemonte.io so the old domain still counts during the move. In DataFast, enable Settings → General → Cookieless / privacy mode so the dashboard matches the script.

There is one custom goal (src/lib/analytics.ts): idea_submitted, with length, saved, and archived params. The idea text is never sent to DataFast.

Evaluation archive (SQLite)

Unless the user checks the opt-out, each successful evaluation creates one row in the evaluations table. It includes the idea, where it came from (source: web or mcp), scoring version, total score, raw verdict, clarity-gate status, category, all eight dimension scores, understandability, latency, model, token usage and the complete diagnostic response JSON. It deliberately excludes IP addresses, cookies and browser identifiers.

For example:

sqlite3 data/analytics.sqlite \
  'SELECT scoring_version, goal, COUNT(*), ROUND(AVG(score), 1) FROM evaluations GROUP BY scoring_version, goal;'

Rows collected before scoring v2 are migrated without being reinterpreted: they remain scoring_version = 1, while v2 stores the new appeal score for fun ideas. To export:

sqlite3 -header -csv data/analytics.sqlite 'SELECT * FROM evaluations;' > evaluations.csv

Scoring benchmark

benchmarks/cases.json contains two similarly detailed, synthetic ideas for each category. The runner evaluates every case against all three goals and always sends doNotArchive: true. With the app running against Jev:

BENCHMARK_RUNS=3 npm run benchmark

It reports score and verdict distributions by goal, category accuracy, clarity gates and repeat-score range. The full JSON is written to the ignored benchmarks/results/latest.json; set BENCHMARK_URL to test a deployed instance.

.mcp.json registers the official DataFast MCP server (https://datafa.st/api/mcp). In Claude Code, run /mcp and sign in with OAuth to query these analytics.

Privacy

• Private history is off by default and stays in the browser's localStorage. There are no accounts.
• Successful evaluations are archived server-side for product analysis by default; the form provides an explicit opt-out.
• DataFast never receives idea text. To move private history to Supabase, implement IdeaStore in src/lib/storage.ts.