Pubcrawl

✨ What is PubCrawl?

PubCrawl connects your AI assistant (Claude Desktop, Cursor, or any MCP-compatible client) directly to the primary sources clinicians and researchers actually use — so you can ask a question in plain English and get an answer grounded in PubMed, Europe PMC, FDA/UK drug labelling, and ClinicalTrials.gov, with real PMIDs, NCT IDs, and DOIs you can verify.

Every tool is a thin, deterministic wrapper over an official API. Nothing is invented; every result cites its source.

The thing no other MCP server does

Ask "compare US and UK labelling for semaglutide" and PubCrawl pulls both live labels and maps equivalent sections — US Indications and Usage ↔ UK 4.1 Therapeutic indications, and so on — so the differences are visible instead of assumed:

[Asking an MCP client to compare US and UK labelling for semaglutide: PubCrawl calls compare_labels, fetches the US Prescribing Information from openFDA/DailyMed and the UK SmPC from the eMC, and returns the two indication lists side by side — three US indications including cardiovascular and renal risk reduction, against one UK indication, with the cardiovascular and renal outcomes appearing only as cross-references to sections 4.4, 4.5 and 5.1.]

A cardiovascular claim that is on-label in the US promotes an unlicensed indication in the UK. If you write, review, or check medical copy for both markets, that gap is the whole job — and compare_labels is the only MCP tool that surfaces it.

Everything else

• 🔬 14 tools across literature, drug labelling, and clinical trials
• 🧾 Verifiable by design — results link back to DailyMed, the eMC, PubMed, and ClinicalTrials.gov
• 📰 Preprints via Europe PMC — surface work ahead of formal publication
• 🆓 No API keys required (an optional free NCBI key raises PubMed rate limits)
• 🧪 Fully typed, tested, and CI-checked — literature retrieval gated by OpenGATE, and labelling & trials by the in-repo fidelity benchmark, on every release

Built by PharmaTools.AI.

🚀 Quick start (60 seconds)

1. Add PubCrawl to your client config — no install step needed, npx fetches it on first run.

For Claude Desktop, edit claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{  "mcpServers": {    "pubcrawl": {      "command": "npx",      "args": ["-y", "@pharmatools/pubcrawl"]    }  }}

2. Restart your client. PubCrawl appears under + → Connectors.

3. Ask away:

"Compare the US and UK labelling for atorvastatin, and find recent Phase 3 trials for it."

That's it. → More examples · API key & other options

🧰 Tools

📚 Literature

💊 Drug labelling

🧫 Clinical trials

💬 Examples

Once connected, just ask naturally:

Literature

• "Search PubMed for recent clinical trials on semaglutide."
• "Search Europe PMC for preprints on GLP-1 receptor agonists, most cited first."
• "Get the abstract for PMID 38127654, then find related papers and cite them all in Vancouver style."
• "What are the trending papers on CRISPR gene therapy this month, high-impact journals only?"
• "Pull the full text of that PMC article and summarise the methods section."

Drug labelling

• "Get the FDA prescribing information for metformin — just the indications and warnings."
• "Pull the UK SmPC for atorvastatin."
• "Compare US and UK labelling for lisinopril and highlight the differences."
• "What's the generic name and drug class for Ozempic?"
• "What drugs are approved for type 2 diabetes in both the US and UK?"

Clinical trials

• "Find recruiting Phase 3 trials for pembrolizumab in breast cancer."
• "Get the eligibility criteria and primary outcomes for NCT03086486."

Cross-source (where PubCrawl shines)

• "For semaglutide: summarise the US label's cardiovascular indication, then find the pivotal trial and its NEJM publication."

🏗 Architecture

Three layers — tools register the MCP interface, lib clients talk to each external API, and shared cache + parsers keep it fast and consistent.

Each tool file exports a register*Tool(server) function with a zod schema and an async handler. All network calls are rate-limited, cached, and time-bounded. See CLAUDE.md for a full architecture walkthrough and CONTRIBUTING.md to add a tool.

🔧 Configuration

Install options

# Zero-install (recommended): npx fetches it on demand — see Quick start above.
# Or install globally:npm install -g @pharmatools/pubcrawl
# Config for a global install:#   { "mcpServers": { "pubcrawl": { "command": "pubcrawl" } } }

NCBI API key (optional)

Without a key, PubMed requests are limited to 3/second. A free key raises this to 10/second.

1. Create a free NCBI account at https://www.ncbi.nlm.nih.gov/account/
2. Account Settings → API Key Management → create a key
3. Add it to your config:

{  "mcpServers": {    "pubcrawl": {      "command": "npx",      "args": ["-y", "@pharmatools/pubcrawl"],      "env": { "NCBI_API_KEY": "your_key_here" }    }  }}

HTTP transport

PubCrawl also ships a stateless Streamable HTTP transport for hosted and browser-based clients:

npm run start:http   # serves POST /mcp and GET /health on PORT (default 3000)

🏢 Run in your own environment

For teams who need PubCrawl inside their own cloud project rather than on a laptop, it runs as a container serving MCP over Streamable HTTP. Each request is handled by a fresh server instance with no sessions, so it scales horizontally behind any load balancer.

docker build -t pubcrawl .docker run --rm -p 8080:8080 pubcrawlcurl http://localhost:8080/health   # {"status":"ok","server":"pubcrawl","version":"…"}

Point any MCP client, or an agent registry such as Google Cloud Agent Registry, Microsoft Foundry or Amazon Bedrock AgentCore, at https://<your-host>/mcp. Authentication is left to the platform in front of the container (Cloud Run IAM, API Gateway, an identity-aware proxy), so PubCrawl itself holds no credentials for your users.

The image runs as a non-root user, has a container health check, and shuts down cleanly on SIGTERM. CI builds it on every push and checks the health endpoint, the MCP handshake, the full tool list and a clean stop.

Google Cloud Run and Agent Registry

These steps were run end to end in November 2026. Replace PROJECT_ID and REGION; the project needs billing and the run, cloudbuild, artifactregistry and agentregistry APIs enabled.

1. Deploy behind Cloud Run IAM. From a clone of this repo (the --functional-type flag needs a current gcloud with the beta component):

gcloud beta run deploy pubcrawl --source . --project=PROJECT_ID --region=REGION \  --no-allow-unauthenticated --functional-type=mcp-server --max-instances=2

Unauthenticated requests to any path now get 403 from Google before reaching PubCrawl. With an identity token, /health answers as usual:

URL=$(gcloud run services describe pubcrawl --project=PROJECT_ID --region=REGION --format='value(status.url)')curl -H "Authorization: Bearer $(gcloud auth print-identity-token)" $URL/health

2. Register the endpoint and its tools. --functional-type=mcp-server adds an entry to Agent Registry that links to the Cloud Run service, but it carries no endpoint URL and no tools, because the registry can't read tools from a server that requires authentication. Register the endpoint manually with the tool spec in deploy/agent-registry/toolspec.json:

gcloud agent-registry services create pubcrawl-mcp --project=PROJECT_ID --location=REGION \  --display-name="PubCrawl" \  --mcp-server-spec-type=tool-spec \  --mcp-server-spec-content=deploy/agent-registry/toolspec.json \  --interfaces=url=$URL/mcp,protocolBinding=jsonrpc

After a minute or so, gcloud agent-registry mcp-servers list --project=PROJECT_ID --location=REGION shows PubCrawl with its /mcp URL and all 14 tools.

The tool spec is generated from the server itself (npm run build && npm run toolspec), and a test fails if the committed file falls out of step with the code. The registry requires an input schema for every tool and caps the spec at 10 KB, so the schemas keep their types, required fields, enums, defaults and ranges but drop parameter descriptions. Clients that connect still get the complete schemas from the server.

3. Connect a client. Any client that can send a header can call the service with a Google identity token. With Claude Code:

claude mcp add --transport http pubcrawl-gcp $URL/mcp \  --header "Authorization: Bearer $(gcloud auth print-identity-token)"

The token expires after an hour and is stored in the client's config, so this suits a quick test; remove it afterwards with claude mcp remove pubcrawl-gcp. For everyday use, gcloud run services proxy pubcrawl --project=PROJECT_ID --region=REGION --port=3000 keeps refreshing your credentials, and the client connects to http://localhost:3000/mcp with no header. The proxy needs the cloud-run-proxy gcloud component. In our test it reported the binary as missing straight after installing it; the header route above works without it.

Agents running in Google Cloud should call the service as their own service account, granted roles/run.invoker on it, rather than with a user's token.

Environment variables

Network egress

PubCrawl needs outbound HTTPS to these hosts only, which is the complete list for a firewall allowlist:

Data handling

• What leaves your environment: the search terms, identifiers and drug names in each tool call are sent to the public source that answers it, from the table above. Nothing goes anywhere else, and PharmaTools.AI receives nothing.
• What is kept: upstream responses are held in an in-memory cache (500 entries, up to 24 hours) to stay within the sources' rate limits. Nothing is written to disk, and the cache is gone when the instance stops.
• What is logged: one line at startup and one at shutdown. Requests and queries are not logged.
• Identification to NCBI: per NCBI's usage policy, E-utilities requests identify the tool as pubcrawl with the maintainer's contact address.
• No model in the path: every tool is a deterministic API call and parse. Nothing is generated.

🗺 Roadmap

Highlights of what's planned — see ROADMAP.md for the full list.

• get_europepmc_fulltext — read preprints & OA articles surfaced by search_europepmc
• get_adverse_events — openFDA FAERS adverse-event lookups
• EMA / EPAR labelling to complement the US + UK compare_labels
• MeSH query helper for sharper PubMed searches
• MCP resources & prompts for common review workflows

Ideas welcome — open an issue.

🛠 Development

git clone https://github.com/nickjlamb/pubcrawl.gitcd pubcrawlnpm install
npm run dev      # TypeScript watch modenpm run build    # compile to dist/npm start        # run the stdio servernpm test         # Vitest unit suitenpm run lint     # ESLint

Unit tests live in tests/ and cover the parsing, caching, citation, formatting and label-pairing logic with fixture payloads (no network calls). CI runs lint → test → build on every push and pull request. New to the codebase? Start with CONTRIBUTING.md.

Measuring fidelity

npm run bench            # labelling + trials fidelity benchmark against the live sourcesnpm run bench -- --ci    # what the release gate runs: exit 1 if a verified case fails

Two gates run on every release and weekly. OpenGATE's retrieval scorer checks PubMed and Europe PMC records against hand-verified anchors. The in-repo fidelity benchmark does the same for drug labelling and trials: compare_labels is held to a gold set of US/UK anchors (a phrase must appear on one side and must not on the other, so real divergences are surfaced rather than smoothed over) plus a verbatim check that refetches the raw openFDA record and eMC page and confirms every returned sentence is a substring of the source. No LLM judge anywhere in either gate.

📦 Releases & changelog

Versions follow Semantic Versioning. See the CHANGELOG for a full history and Releases for notes and assets.

🤝 Contributing

Contributions are welcome and appreciated — bug reports, new data sources, new tools. Read the contributing guide to get started, then open an issue or a pull request.

📚 Citation

If PubCrawl supports work you publish, please cite it — see CITATION.cff, or:

@software{lamb_pubcrawl,  author    = {Lamb, Nick},  title     = {PubCrawl: verifiable biomedical literature, drug labelling and clinical trials for AI assistants},  year      = {2026},  publisher = {Zenodo},  doi       = {10.5281/zenodo.22101559},  url       = {https://doi.org/10.5281/zenodo.22101559}}

📄 License

MIT © PharmaTools.AI