ToolRank (private MVP → public-ready deploy)

Ranked shortlist of MCP servers / APIs for agent jobs — health, price_band, and install snippet.

Live: https://toolrank.onrender.com — OpenAPI at /openapi.json, Streamable HTTP MCP at /mcp.

Private build for Brandon Clark / ToolRank. All Rights Reserved (see LICENSE). GitHub: brandonclark1996-design/toolrank.

Honesty: Ranking is lexical (token overlap), not learned or sponsored. Catalog is 36 curated tools. health is often unknown. Stripe is not wired. Custom domain is not purchased.

Stack

• Node 20, TypeScript

• Hono HTTP API

• Zod validation

• JSON catalog (data/catalog.json, 36 curated public tools)

• MCP stdio (npm run mcp) and Streamable HTTP (POST /mcp) via @modelcontextprotocol/sdk

• API-key auth + process-local soft quotas (src/auth.ts) on /v1/* and /mcp

Related MCP server: disvr

Quick start (local)

cd /workspace/toolrank
cp .env.example .env
# Edit .env — set TOOLRANK_API_KEYS if you want keyed access locally
export TOOLRANK_API_KEYS='trk_live_test:builder'
npm install
npm test
npm run build
npm start              # listens on 0.0.0.0:$PORT (default 8787)
# or: npm run dev

Local curl with a test key

export TOOLRANK_API_KEYS='trk_live_test:builder'
npm start

curl -s http://127.0.0.1:8787/health | jq .

curl -s http://127.0.0.1:8787/v1/search \
  -H 'content-type: application/json' \
  -H 'Authorization: Bearer trk_live_test' \
  -d '{"job":"scrape website to markdown for RAG","constraints":{"budget":"any"},"limit":5}' | jq .

# Or X-API-Key:
curl -s http://127.0.0.1:8787/v1/search \
  -H 'content-type: application/json' \
  -H 'X-API-Key: trk_live_test' \
  -d '{"job":"web search for research","limit":3}' | jq .

Anonymous (no key) works until the free IP quota is hit (20 / rolling 24h).

Auth + limits

• Env: TOOLRANK_API_KEYS = comma-separated keys, optional key:plan (free | builder). Example: trk_live_abc:builder,trk_live_xyz:free. Bare key → free.

• /health and /openapi.json are always open (no auth / no quota).

• Invalid key → 401 { error, schema_version, as_of }

• Over quota → 429 + Retry-After (seconds hint)

• Authed POST /v1/search responses include plan (free | builder)

• Quotas are in-memory / process-local — fine for single-instance MVP; they reset on restart and do not sync across replicas.

Setting keys on the host (Brandon)

Do not commit real keys. Set TOOLRANK_API_KEYS in the host environment:

1. Render: Dashboard → your Web Service → Environment → add TOOLRANK_API_KEYS = trk_live_….…:builder,… (use sync: false / secret). Redeploy.

2. Railway: Service → Variables → TOOLRANK_API_KEYS.

3. Docker / VPS: -e TOOLRANK_API_KEYS='…' or systemd/env file outside git.

4. Local: export TOOLRANK_API_KEYS='trk_live_test:builder' before npm start.

Also set HOST=0.0.0.0 (default in code) and let the platform inject PORT.

HTTP API

Base: https://toolrank.onrender.com (or http://127.0.0.1:8787 locally).

Search body

{
  "job": "string (required)",
  "constraints": {
    "budget": "free | paid | any",
    "auth": "optional hint",
    "runtime": "mcp | api | …"
  },
  "limit": 10,
  "check_health": false
}

Optional check_health: true probes top results (2s timeout). Failures stay unknown (no false down).

Deploy (no custom domain / no Stripe)

Docker

docker build -t toolrank .
docker run --rm -p 8787:8787 \
  -e TOOLRANK_API_KEYS='trk_live_prod:builder' \
  toolrank

Render

• Blueprint: render.yaml (Docker web service, free plan friendly).

• Live hostname: https://toolrank.onrender.com (no custom domain required).

• Set TOOLRANK_API_KEYS in the dashboard.

Railway

• railway.toml (Dockerfile) or Procfile (web: node dist/index.js after build).

• Set TOOLRANK_API_KEYS in Variables. Use Railway’s public HTTP domain.

Build step for native Node hosts: npm ci && npm run build then node dist/index.js.

MCP

Tools: search_tools, get_tool, list_fresh (shared handlers for stdio + HTTP).

Streamable HTTP (production)

https://toolrank.onrender.com/mcp

Same API-key / anonymous quota as /v1/*. Official MCP Registry manifest: server.json (io.github.brandonclark1996-design/toolrank). Do not run mcp-publisher from this box unless asked.

Cursor / Claude HTTP MCP config snippet:

{
  "mcpServers": {
    "toolrank": {
      "url": "https://toolrank.onrender.com/mcp"
    }
  }
}

stdio (local)

npm run mcp

{
  "mcpServers": {
    "toolrank": {
      "command": "npm",
      "args": ["run", "mcp"],
      "cwd": "/workspace/toolrank"
    }
  }
}

Ranking

1. Tokenize job (drop stopwords).

2. Score tag / name / description matches.

3. Boost matching price_band under budget constraints.

4. Boost health: up; penalize health: down.

5. Tiny preference for mcp / both interfaces.

Honest lexical ranking — no paid placement. 36 curated tools. Health often unknown. Stripe not wired.

Docs

• docs/AGENT.md — agent-readable usage (auth headers + limits)

• docs/llms.txt — compact machine summary

• server.json — Official MCP Registry manifest (not published yet)

• scripts/README-ingest.md — stub for future registry.modelcontextprotocol.io ingest

Tests & build

npm test    # vitest — ranking, envelope, auth middleware, /mcp
npm run build

TODO

• Implement registry ingest (scripts/README-ingest.md)

• Persist health probe results between restarts

• Stripe / paid upgrades (not started)

• Custom domain (not purchased / not wired)

• Shared Redis (or similar) quotas if multi-instance

• Publish server.json with mcp-publisher (not run from this change)

License

All Rights Reserved — Brandon Clark / ToolRank. See LICENSE.