# BioMCP

One binary. One grammar. Evidence from the biomedical sources you already trust.

## Data terms

BioMCP retrieves data from upstream biomedical sources. Those sources' terms govern how you use the data they return: some restrict commercial or clinical use. Check each source's terms and obtain any licence you need before relying on the data. See [Source Licensing and Terms](docs/reference/source-licensing.md) for the per-source breakdown.

## What is BioMCP?

BioMCP is one CLI binary over a single command grammar that reaches 72 trusted
biomedical sources directly (PubMed, ClinVar, ClinicalTrials.gov, OncoKB,
Reactome, and more; ten further sources arrive inside another source's
answer). It is also an MCP (Model Context Protocol) server, so the same tools are
available to AI agents such as Claude Code, Codex, and Claude Desktop.

BioMCP cuts through the usual biomedical data maze: one query reaches the
sources that normally live behind different APIs, identifiers, and search
habits. Researchers, clinicians, and agents use the same command grammar to
search, focus, and pivot without rebuilding the workflow for each source. You
get compact, evidence-oriented results across live public data plus local study analytics.

## Features

- **Search the literature:** `search article` fans out across PubTator3 and
  Europe PMC and PubMed when the filters are compatible, deduplicates PMID/PMCID/DOI identifiers, and can add a Semantic
  Scholar leg when your filters support it.
- **Pivot without rework:** move from a gene, variant, drug, disease, pathway,
  protein, or article straight into the next built-in view instead of
  rebuilding filters by hand.
- **Choose a playbook:** `biomcp skill list` shows shipped worked examples
  so you can open the matching `biomcp skill <slug>` workflow.
- **Analyze studies locally:** `study` commands cover local query, cohort, survival,
  compare, and co-occurrence workflows with native terminal, SVG, and PNG
  charts for downloaded cBioPortal-style datasets.
- **Follow the paper trail:** `article citations`, `article references`,
  `article recommendations`, and `article entities` turn one known paper into a
  broader evidence map.
- **Enrich and batch:** use `biomcp enrich` for top-level g:Profiler
  enrichment and `biomcp batch` for up to 10 focused `get` calls in one
  command.

## Quick start

First useful query in under 30 seconds:

```bash
uv tool install biomcp-cli
biomcp health --apis-only
biomcp skill list
biomcp list gene
biomcp search all --gene BRAF --disease melanoma  # unified cross-entity discovery
biomcp get gene BRAF pathways hpa
```

## Watch the introduction

A project using BioMCP won St. Jude's KIDS BioHackathon in 2025. The talk that
followed is a one-hour introduction to BioMCP, recorded October 24, 2025.

[![BioMCP: An Introduction to Biomedical AI Agents](docs/assets/st-jude-2025-talk-1280x720.png)](https://www.youtube.com/watch?v=lXoe-4TENDE)

[Watch on YouTube](https://www.youtube.com/watch?v=lXoe-4TENDE) &middot;
[Read the write-up](https://biomcp.org/blog/st-jude-2025-hackathon-talk/) &middot;
[Full transcript](https://www.imaurer.com/talks/biomcp-st-jude/)

## Installation

### Binary install

```bash
curl -fsSL https://biomcp.org/install.sh | bash
```

### PyPI tool install

```bash
uv tool install biomcp-cli
# or: pip install biomcp-cli
```

> **PyPI package warning:** install `biomcp-cli`, not `biomcp`. The `biomcp`
> PyPI package is unrelated to this project.

MCP Registry ownership marker: `mcp-name: io.github.genomoncology/biomcp`.

This installs the `biomcp` binary in `~/.local/bin`. If that directory is not
already on `PATH`, the installer prints one command to add it; it never edits
your shell startup files.

### Homebrew

```bash
brew tap genomoncology/biomcp
brew install biomcp
```

The separate `genomoncology/homebrew-biomcp` tap repository must exist before these commands can work.

### Docker

```bash
docker run --rm ghcr.io/genomoncology/biomcp --version
docker run --rm ghcr.io/genomoncology/biomcp list
docker run --rm -i ghcr.io/genomoncology/biomcp serve
```

Use the GHCR image for quick CLI checks or stdio MCP clients without a local install.

### Claude Code plugin

Install the `biomcp` binary first, then add the hosted plugin marketplace and
install the BioMCP plugin in Claude Code:

```text
/plugin marketplace add genomoncology/biomcp
/plugin install biomcp@biomcp
```

The plugin wires Claude Code to the local stdio MCP server with `biomcp serve`.
For guided BioMCP workflows, also install the skill assets below.

### Codex MCP server

Install the `biomcp` binary first, then register the same stdio MCP server with
Codex:

```bash
codex mcp add biomcp -- biomcp serve
```

### Claude Desktop extension (.mcpb)

Install BioMCP from the Anthropic Directory in Claude Desktop when that path is
available for your environment. For local/manual setups, use the JSON MCP
config below.

### Install skills

The executable embeds the guide and worked examples. Read the complete guide
with `biomcp skill render`, or install copies into your agent directory:

```bash
biomcp skill install ~/.claude --force
```

### MCP clients

```json
{
  "mcpServers": {
    "biomcp": {
      "command": "biomcp",
      "args": ["serve"]
    }
  }
}
```

### Remote HTTP server

For shared or remote deployments:

```bash
biomcp serve-http --host 127.0.0.1 --port 8080
```

Remote clients connect to `http://127.0.0.1:8080/mcp`. Probe routes are
`GET /health`, `GET /readyz`, and `GET /`.

Runnable demo:

```bash
uv run --script examples/streamable-http/streamable_http_client.py
```

See [Remote HTTP Server](https://biomcp.org/getting-started/remote-http/) for
the newcomer guide.

### From source

```bash
make install
"$HOME/.local/bin/biomcp" --version
```

For repo-local verification, run the standard gates directly: `make lint`,
`make test`, and `make spec`. `make test` includes both Rust nextest and the
Python/docs contract lane, while `make release-gate` adds the named full-feature
proof and runs specs against the all-feature release binary.
There is no supported `make check` command. Use `make verify` only for opt-in
live public-upstream confidence; `make release-live-smoke` remains a
compatibility alias.

## Command grammar

```text
search <entity> [filters]    → discovery
skill list                   → playbook catalog for how-to questions
discover <query>             → concept resolution before entity selection
get <entity> <id> [sections] → focused detail
<entity> <helper> <id>       → cross-entity pivots
enrich <GENE1,GENE2,...>     → gene-set enrichment
batch <entity> <id1,id2,...> → parallel gets
search all [slot filters]    → counts-first cross-entity orientation
```

## Entities and sources

The tables below distinguish detail-card entities from search-only surfaces so
agents do not synthesize unsupported `get` commands.

### Gettable entities

| Entity | Upstream providers used by BioMCP | Example |
|--------|-----------------------------------|---------|
| gene | MyGene.info, UniProt, Reactome, QuickGO, STRING, GTEx, Human Protein Atlas, DGIdb, ClinGen, NIH Reporter, DisGeNET, GTR-backed diagnostics pivot | `biomcp get gene BRAF pathways hpa` |
| variant | MyVariant.info, ClinVar, direct gnomAD v4 population data, CIViC, Cancer Genome Interpreter, OncoKB, cBioPortal, GWAS Catalog, AlphaGenome | `biomcp get variant "BRAF V600E" clinvar` |
| article | PubMed, PubTator3, Europe PMC, PMC OA, NCBI ID Converter, Semantic Scholar (optional auth; `S2_API_KEY` recommended) | `biomcp search article -g BRAF --limit 5` |
| author | Semantic Scholar, ORCID | `biomcp get author semanticscholar:1716151` |
| trial | ClinicalTrials.gov API v2, NCI CTS API | `biomcp search trial -c melanoma -s recruiting` |
| diagnostic | NCBI Genetic Testing Registry local bulk bundle + WHO IVD local CSV + optional OpenFDA device overlay | `biomcp get diagnostic GTR000006692.3 regulatory` |
| drug | MyChem.info, DDInter local bundle, EMA local batch, WHO Prequalification local exports, ChEMBL, OpenTargets, Drugs@FDA, OpenFDA labels/shortages/approvals/FAERS/MAUDE/recalls, CIViC | `biomcp drug interactions warfarin` |
| disease | MyDisease.info, Monarch Initiative, MONDO, OpenTargets, Reactome, CIViC, SEER Explorer, NIH Reporter, DisGeNET, GTR/WHO IVD diagnostics pivot | `biomcp get disease "Lynch syndrome" genes` |
| pathway | Reactome, KEGG, WikiPathways, g:Profiler, Enrichr-backed enrichment sections | `biomcp get pathway hsa05200 genes` |
| protein | UniProt, InterPro, STRING, ComplexPortal, PDB, AlphaFold | `biomcp get protein P15056 complexes` |
| adverse-event | OpenFDA FAERS/MAUDE/recalls plus CDC WONDER VAERS aggregate vaccine search | `biomcp search adverse-event --drug pembrolizumab` |
| pgx | CPIC, PharmGKB | `biomcp get pgx CYP2D6 recommendations` |
| cell-line | Cellosaurus with ChEMBL, PharmacoDB, DepMap, COSMIC, GDSC, and LINCS cross-references | `biomcp get cell-line CVCL_0004 xrefs` |

### Search-only entities

| Entity | Upstream providers used by BioMCP | Example |
|--------|-----------------------------------|---------|
| gwas | GWAS Catalog | `biomcp search gwas --trait "type 2 diabetes"` |
| phenotype | Monarch Initiative (HPO semantic similarity plus direct-support checks) | `biomcp search phenotype "HP:0001250"` |

## Cross-entity helpers

Pivot between related entities without rebuilding filters.

See the [cross-entity pivot guide](docs/how-to/cross-entity-pivots.md) for when
to use a helper versus a fresh search.

```bash
biomcp variant trials "BRAF V600E" --limit 5
biomcp variant articles "BRAF V600E"
biomcp drug adverse-events pembrolizumab
biomcp drug trials pembrolizumab
biomcp disease trials melanoma
biomcp disease drugs melanoma
biomcp disease articles "Lynch syndrome"
biomcp gene trials BRAF
biomcp gene drugs BRAF
biomcp gene articles BRCA1
biomcp gene pathways BRAF
biomcp pathway drugs R-HSA-5673001
biomcp pathway drugs hsa05200
biomcp pathway articles R-HSA-5673001
biomcp pathway trials R-HSA-5673001
biomcp protein structures P15056
biomcp article entities 22663011
biomcp article citations 22663011 --limit 3
biomcp article references 22663011 --limit 3
biomcp article recommendations 22663011 --limit 3
```

## Gene-set enrichment

```bash
biomcp enrich BRAF,KRAS,NRAS --limit 10
```

Top-level `biomcp enrich` uses **g:Profiler**. Gene enrichment sections inside
other entity views still reference **Enrichr** where that is the backing
source.

## Sections and progressive disclosure

Every `get` command supports selectable sections for focused output:

```bash
biomcp get gene BRAF                    # summary card
biomcp get gene BRAF pathways           # add pathway section
biomcp get gene BRCA1 diagnostics       # diagnostic-test pivot from GTR
biomcp get gene BRAF hpa                # protein tissue expression + localization
biomcp get gene BRAF civic interactions # multiple sections
biomcp get gene BRAF all                # standard sections; diagnostics/funding stay opt-in

biomcp get variant "BRAF V600E" clinvar population conservation
biomcp get article 22663011 tldr
biomcp get drug pembrolizumab label targets civic approvals
biomcp get drug trastuzumab regulatory --region who
biomcp get disease "Lynch syndrome" genes phenotypes variants
biomcp get disease tuberculosis diagnostics
biomcp get diagnostic GTR000006692.3 regulatory
biomcp get trial NCT02576665 eligibility locations outcomes
```

In JSON mode, `get` responses expose `_meta.next_commands` for the next likely
follow-ups and `_meta.section_sources` for section-level provenance. `batch ...
--json` returns per-entity objects with the same metadata shape.

## API keys

Most commands work without credentials. Optional keys improve rate limits or
unlock optional enrichments:

```bash
export NCBI_API_KEY="..."        # ClinVar, PubTator, PubMed/efetch, PMC OA, NCBI ID converter
export S2_API_KEY="..."          # Optional Semantic Scholar auth; dedicated quota at 1 req/sec
export OPENFDA_API_KEY="..."     # OpenFDA rate limits
export NCI_API_KEY="..."         # NCI CTS trial search (--source nci)
export ONCOKB_TOKEN="..."        # OncoKB variant helper
export ALPHAGENOME_API_KEY="..." # AlphaGenome variant effect prediction
```

`search article`, `get article`, `batch article`, `get article ... tldr`, and
the explicit Semantic Scholar helpers all work without `S2_API_KEY`. With the
key, BioMCP sends authenticated requests and uses a dedicated rate limit at
1 req/sec. Without it, BioMCP uses the shared unauthenticated pool at 1 req/2sec.
`search article --source` supports `all`, `pubtator`, `europepmc`, `pubmed`,
`semanticscholar`, and `litsense2`. The default compatible article federation
uses PubTator3, Europe PMC, PubMed, and automatic Semantic Scholar; use
`--source semanticscholar` or `--source litsense2` explicitly when you want one
of those sources alone. Explicit source selection also disables cross-provider
row enrichment. References
and recommendations can be empty for paywalled papers because of publisher
elision in Semantic Scholar upstream coverage.

## Configuration

### Claude Desktop extension settings

The directory bundle exposes only the optional settings needed for the first
reviewer-facing build:

| Claude Desktop field | Runtime env var | Purpose |
|----------------------|-----------------|---------|
| OncoKB Token | `ONCOKB_TOKEN` | Enables `biomcp variant oncokb "<gene> <variant>"` therapy and level evidence |
| DisGeNET API Key | `DISGENET_API_KEY` | Enables scored DisGeNET sections on gene and disease lookups |
| Semantic Scholar API Key | `S2_API_KEY` | Improves reliability for article TLDR, citation, reference, and recommendation helpers |

The first directory build exposes only those three optional settings. Advanced
CLI-only env vars remain documented in
[API Keys](docs/getting-started/api-keys.md) for the general BioMCP CLI path.

## Usage Examples

### Public cross-entity overview

**User prompt:** Give me a low-noise overview of BRAF in melanoma.

**Expected tool call:** `biomcp search all --gene BRAF --disease melanoma --counts-only`

**Expected behavior:** Returns a cross-entity counts summary that orients the
next command instead of dumping long detail tables.

**Expected output:** Counts-first summary with suggested next commands for the
highest-yield entity follow-ups.

### Public variant evidence

**User prompt:** Summarize ClinVar significance and population frequency for BRAF V600E.

**Expected tool call:** `biomcp get variant "BRAF V600E" clinvar population`

**Expected behavior:** Retrieves the focused variant card, ClinVar section, and
population-frequency data in one read-only call.

**Expected output:** Variant summary, ClinVar significance details, and gnomAD
population frequencies.

### Credentialed OncoKB example

**User prompt:** Show OncoKB therapy evidence for BRAF V600E.

**Expected tool call:** `biomcp variant oncokb "BRAF V600E"`

**Expected behavior:** Uses `ONCOKB_TOKEN` when configured and otherwise
returns helpful guidance about the missing credential.

**Expected output:** Therapy and level evidence when `ONCOKB_TOKEN` is set, or
a clear setup hint when it is not.

### Credentialed DisGeNET example

**User prompt:** Show scored DisGeNET associations for TP53.

**Expected tool call:** `biomcp get gene TP53 disgenet`

**Expected behavior:** Uses `DISGENET_API_KEY` to retrieve the scored
gene-disease association section.

**Expected output:** Ranked disease-association table with evidence counts and
scores when `DISGENET_API_KEY` is configured.

## Privacy Policy

BioMCP does not add telemetry, analytics, or remote log upload. Review the
full privacy statement at <https://biomcp.org/policies/>.

## Multi-worker deployment

BioMCP rate limiting is process-local. For many concurrent workers, run one shared
Streamable HTTP `biomcp serve-http` endpoint so all workers share a single
limiter budget:

```bash
biomcp serve-http --host 0.0.0.0 --port 8080 \
  --allowed-hosts biomcp.example.org
```

Loopback servers accept only local `Host` values by default. A non-loopback
bind requires `--allowed-hosts`. The explicit `--unsafe-allow-any-host` escape
hatch disables only this Host check; it does not add authentication, TLS, or
encryption. Put remote deployments behind a trusted authenticated TLS proxy or
inside a private network.

Remote clients should connect to `http://<host>:8080/mcp`. Lightweight process
probes are available at `GET /health`, `GET /readyz`, and `GET /`.

## Skills

BioMCP ships an embedded agent guide and a worked-example catalog.
Use `biomcp skill list` when you need the right worked example,
then use `biomcp skill` to read the embedded BioMCP guide or install it into
your agent directory when you want local copies of the workflow references:

```bash
biomcp skill list
biomcp skill
biomcp skill install ~/.claude --force
```

See [Skills](docs/getting-started/skills.md) for supported install targets,
installed files, and legacy compatibility notes.

## Local study analytics

`study` is BioMCP's local analysis family for downloaded cBioPortal-style datasets.
The public entity surface handles API-backed, local-runtime, and hybrid
discovery/detail; `study` commands work on local datasets when you need
per-study query, cohort, survival, comparison, or co-occurrence workflows. Per-gene queries include mutations, CNA, expression, and structural variants/fusions from local `data_sv.txt` files. Mutation summaries stay mutation-only and note when fusions/SV need `--type sv`.

Use `study download` to fetch a dataset into your local study root. Set
`BIOMCP_STUDY_DIR` when you want an explicit dataset location for reproducible
scripts and demos; if it is unset, BioMCP falls back to its default study root.

```bash
export BIOMCP_STUDY_DIR="$HOME/.local/share/biomcp/studies"
biomcp study download msk_impact_2017
biomcp study query --study msk_impact_2017 --gene TP53 --type mutations --chart bar --theme dark --palette wong -o docs/blog/images/tp53-mutation-bar.svg
biomcp study query --study msk_impact_2017 --gene RET --type sv
```

See the [CLI reference](docs/user-guide/cli-reference.md#local-study-analytics)
for the full `study` command family and dataset prerequisites.

## Ops

```bash
biomcp version                            # show version and build info
biomcp health                             # inspect API connectivity plus local DDInter/EMA/cache readiness
biomcp update                             # self-update with release SHA256 checksum verification
biomcp update --check                     # check for updates without installing
biomcp uninstall                          # remove biomcp from ~/.local/bin
```

## Support

- GitHub issues: <https://github.com/genomoncology/biomcp/issues>
- Troubleshooting: [docs/troubleshooting.md](docs/troubleshooting.md)
- Full documentation: <https://biomcp.org/>

## Documentation

- [Getting Started](docs/getting-started/installation.md)
- [Search All Workflow](docs/how-to/search-all-workflow.md)
- [BioASQ Benchmark](docs/reference/bioasq-benchmark.md)
- [Cross-Entity Pivot Guide](docs/how-to/cross-entity-pivots.md)
- [Privacy Policy](docs/policies.md)
- [Source Licensing and Terms](docs/reference/source-licensing.md)
- [Data Sources](docs/reference/data-sources.md)
- [Quick Reference](docs/reference/quick-reference.md)
- [Troubleshooting](docs/troubleshooting.md)

## Citation

If you use BioMCP in research, cite it via [`CITATION.cff`](CITATION.cff).
GitHub also exposes `Cite this repository` in the repository sidebar when that file is present.

## Data Sources and Licensing

BioMCP is MIT-licensed. It performs on-demand queries against upstream providers instead of vendoring or mirroring their datasets, but upstream terms govern reuse of retrieved results.

Some providers are fully open, some BioMCP features require registration or API keys, and some queryable sources still impose notable reuse limits. The two biggest cautions are KEGG, which distinguishes academic and non-academic use, and COSMIC, which BioMCP keeps indirect-only because its licensing model is incompatible with a direct open integration.

Use [Source Licensing and Terms](docs/reference/source-licensing.md) for the per-source breakdown and [API Keys](docs/getting-started/api-keys.md) for setup steps and registration links.

## License

MIT