BioHarbor README

BioHarbor

Run real bioinformatics from your AI agent. Reliable, reproducible, on your own GPUs.

🧪 Alpha (v0.1). Sequence tools, homology search (MMseqs2) and structure prediction (ESMFold, validated on RTX 5090) work. Feedback welcome — see the roadmap.

BioHarbor is an MCP server that lets AI agents such as Claude, Cursor and Codex execute bioinformatics tools — not just look things up. Agents ask for an analysis; BioHarbor validates the input, schedules it on a GPU with room, records exactly how it ran, and hands back a compact, agent-readable summary.

BioHarbor demo

Why another bio MCP server?

Most bio MCP servers wrap databases (UniProt, PDB, PubMed…). Use them — BioHarbor complements them by running the compute:

	Database MCP servers	BioHarbor
Runs real analyses (search, fold, cluster)	❌	✅
Validates inputs before burning GPU time	❌	✅
GPU-aware queue, polite on shared GPUs	❌	✅
Long jobs return a `job_id` instead of timing out	❌	✅
Compact summaries + files on disk (saves tokens)	❌	✅
Provenance for every run, export to a pipeline	❌	✅ (export: planned)

Quick start

Terminal

pip install bioharbor
bioharbor doctor             # checks Python, GPUs, workspace, tools
bioharbor setup-db swissprot # reference database for search_homologs (needs MMseqs2)
bioharbor install            # shows how to connect Claude, Cursor or Codex

For structure prediction on a GPU: pip install "bioharbor[esmfold]" — see docs/gpu-setup.md (RTX 50xx needs a CUDA 12.8+ PyTorch).

Connect your agent

BioHarbor is a standard MCP server, so it works with any MCP client. One command sets up the popular ones (it writes an absolute path, so GUI apps find it even outside your venv):

Client	Set up
Claude Code	`claude mcp add bioharbor -- bioharbor serve`
Claude Desktop	`bioharbor install claude-desktop --write`, then restart the app
Cursor	`bioharbor install cursor --write`, or
Codex (CLI, IDE extension, app)	`codex mcp add bioharbor -- bioharbor serve`, or `bioharbor install codex --write`
Biomni (Stanford's biomedical agent)	`agent.add_mcp(...)`; see docs/use-with-biomni.md
Anything else	run `bioharbor serve` (stdio) or `bioharbor serve --http` (Streamable HTTP)

Long-running tools return a job_id within ~20 s instead of blocking, so they stay within every client's tool-call timeout.

Codex chaining BioHarbor's find_orfs and seq_stats tools on a DNA sequence

Step-by-step setup (local or on a GPU server, with troubleshooting): docs/connect-clients.md.

Then ask your agent something like:

Find the longest ORF in this contig, translate it, search Swiss-Prot for homologs and predict its structure. Which regions are low confidence?

Use it without an agent

Every tool is also a CLI command, with identical behaviour:

bash

bioharbor tools list
bioharbor run find_orfs sequence=@contig.fa min_aa=100 --brief   # human-readable
bioharbor run seq_stats sequence=MKTAYIAKQRQISFVKSHFSRQ
bioharbor jobs

Shared GPU server

GPUs on a lab server, agent on your laptop? Run BioHarbor on the server and reach it through an SSH tunnel; no extra port is opened on the server:

bash

# on the GPU server
bioharbor serve --http --host 127.0.0.1 --port 8765
# on your laptop, then point Cursor / Codex / Claude Code at http://127.0.0.1:8765/mcp
ssh -N -L 8765:127.0.0.1:8765 you@gpu-server

⚠️ HTTP mode has no authentication yet (on the roadmap), so keep it on 127.0.0.1 and use the tunnel. Details: docs/connect-clients.md.

Tools

Tool	What it does	Runs
`seq_stats`	Validate sequences; type, length, GC%, molecular weight	inline
`translate_sequence`	DNA/RNA → protein, one or all six frames	inline
`find_orfs`	Longest ORFs on both strands, with coordinates	inline
`search_homologs`	MMseqs2 search (protein, or translated DNA) vs local DBs	job
`predict_structure`	ESMFold structure, pLDDT bands, low-confidence regions, pTM	job (GPU)
`scrna_pipeline`	scanpy QC → clustering → markers	planned

Runtime tools: get_job, list_jobs, cancel_job, describe_tool, list_databases, gpu_status, read_file.

How it works

Code

Agent ──MCP──▶ validate input ─▶ inline? ──yes──▶ run ─┐
                                   │ no                  ├─▶ provenance + summary ─▶ Agent
                                   ▼                     │
                     job queue (SQLite) ─▶ GPU placement ┘
                     (waits politely for a GPU with free memory)

Every call is a job recorded in SQLite with params, versions, timings and GPU used, plus a provenance.json next to its outputs.
GPU placement reads live free memory and utilisation (NVML or nvidia-smi), keeps headroom, and reserves memory for jobs it has started so two jobs never grab the same space. Other users' processes are respected.
Fail fast: input, binaries and databases are checked before a job is queued, so a bad request never waits behind a busy GPU.
Results are agent-shaped: summary, message, files, suggestions. Errors carry a hint and a retryable flag.

Details: docs/design.md.

Writing a tool

server.ts

from pydantic import BaseModel, Field
from bioharbor.registry import Resources, RunContext, tool
from bioharbor.results import ToolResult


class FoldParams(BaseModel):
    sequence: str = Field(..., description="Protein sequence")


@tool(
    version="1",
    slow=True,
    resources=Resources(gpu=True, gpu_mem_gb=lambda p: 4 + len(p.sequence) / 100),
)
def predict_structure(params: FoldParams, ctx: RunContext) -> ToolResult:
    """Predict a protein structure with ESMFold."""
    ...
    return ToolResult(summary={"mean_plddt": 87.1}, files=["model.pdb"])

Plugins can ship tools in their own package via the bioharbor.tools entry-point group. See CONTRIBUTING.md.

License

Apache-2.0