deep-research

Give it a company and a question. Get a specific answer.

Costs ~$0.0003–$0.0005 per run. No browser automation — plain HTTP fetch only.

Use in Claude Code (agents dialog)

One command — no cloning, no setup:

claude mcp add deep-research -e OPENROUTER_API_KEY=sk-or-... -- npx deep-research-mcp

Get an OpenRouter key at openrouter.ai (free to sign up, ~$0.0003/run).

For multi-domain search, add your Serper key too:

claude mcp add deep-research -e OPENROUTER_API_KEY=sk-or-... -e SERPER_API_KEY=... -- npx deep-research-mcp

Then in any Claude Code session just ask naturally:

"Research what equipment hatchstamping.com operates"

Claude calls the research tool automatically and returns the answer.

Use in Clay

1. Deploy to Railway (see below) — takes 3 minutes
2. In Clay, add an HTTP API enrichment column
3. POST https://your-app.railway.app/research
4. Body: { "query": "what does {{domain}} sell?", "mode": "single" }
5. Map answer and prospectFit.score to columns

Deploy to Railway

1. Fork this repo on GitHub
2. Go to railway.app → New Project → Deploy from GitHub repo → select your fork
3. Add env vars: OPENROUTER_API_KEY (required), SERPER_API_KEY (optional)
4. Deploy — Railway auto-detects Node and runs npm start, gives you a public URL

Quick start

git clone https://github.com/yourusername/deep-research
cd deep-research
npm install
cp .env.example .env
# add your OPENROUTER_API_KEY to .env
npm start

curl -X POST http://localhost:3000/research \
  -H "Content-Type: application/json" \
  -d '{"query": "what equipment does hatchstamping.com operate at their facilities?"}'

{
  "answer": "Hatch Stamping operates stamping presses, robotic welding cells, progressive die tooling, 3D printing, laser scanning, and CMMs across their facilities.",
  "prospectFit": {
    "score": "high",
    "reason": "Manufacturer that designs and maintains its own heavy equipment."
  },
  "sources": ["https://hatchstamping.com", "https://hatchstamping.com/capabilities"],
  "costUSD": 0.000398,
  "model": "google/gemini-2.0-flash-lite-001",
  "mode": "single"
}

Two modes

Single search (default)

Crawls the company's own website using a ReAct loop. Include a domain in your query — no startUrl needed.

{
  "query": "what equipment does hatchstamping.com operate?",
  "mode": "single"
}

Multi search

Searches across the web via Serper.dev. Better for companies with thin websites or when you need LinkedIn/news/third-party coverage. Requires SERPER_API_KEY.

{
  "query": "what equipment does Hatch Stamping operate?",
  "mode": "multi"
}

API

POST /research

| Field | Type | Required | Description | |-------|------|----------|-------------| | query | string | yes | Your research question. Include the company domain for single mode. | | mode | string | no | single (default) or multi | | model | string | no | OpenRouter model ID (default: google/gemini-2.0-flash-lite-001) | | startUrl | string | no | Override the URL for single mode instead of extracting from query |

Response

| Field | Description | |-------|-------------| | answer | Specific answer with names, numbers, dates | | prospectFit.score | high / medium / low / wrong_target | | prospectFit.reason | One sentence about the company | | prospectFit.angle | Best approach if fit is not high | | sources | Pages visited (single) or search results fetched (multi) | | steps | Each navigation step with findings and specificity rating (single only) | | costUSD | LLM cost only. Add $0.001 per run for Serper if using multi. | | mode | single or multi |

Environment variables

| Variable | Required | Description | |----------|----------|-------------| | OPENROUTER_API_KEY | yes | Get one at openrouter.ai | | SERPER_API_KEY | only for multi | Get one at serper.dev |

Models

Any model on OpenRouter. Recommended:

| Model | Cost/run | Notes | |-------|----------|-------| | google/gemini-2.0-flash-lite-001 | ~$0.0003 | Default. Best cost/quality. | | google/gemini-2.5-flash-lite | ~$0.0004 | Slightly better reasoning | | qwen/qwen3-4b:free | $0 | Free but rate limited |

How it works

Single mode — ReAct loop:

1. Fetches homepage, extracts internal nav links
2. Step 0 builds company context (industry, type, vocabulary)
3. Each step: reads current page → rates specificity → picks next path or declares done
4. Information density filtering removes marketing fluff before the LLM sees anything
5. Two-stage synthesis: compress findings → generate final answer + fit score

Multi mode — Web search:

1. Fires one Serper search with your query
2. Fetches top 3 result pages
3. Runs the same density filter + synthesis

See docs/cost-reduction.md for how cost dropped from $0.11 → $0.0003.

Limitations

• JS-rendered sites — SPAs that return no nav links will get weak answers in single mode. Use multi mode for these.
• Single mode is domain-scoped — only crawls the company's own website.
• Public pages only — no login support.

License

Apache 2.0