# BioHarbor [Health: Active]

**Category:** 🧬 Biology & Bioinformatics  
**Repository:** https://github.com/danielLuo2/bioharbor  
**GitHub Stars:** 0  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/bioharbor

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

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "bioharbor": {
    "command": "uvx",
    "args": ["bioharbor"]
  }
}
```

## Documentation

## What BioHarbor MCP server does

BioHarbor MCP server exposes local bioinformatics operations to MCP-compatible agents. Its current working tools cover sequence statistics, sequence translation, open reading frame discovery, MMseqs2 homology searches, and ESMFold protein structure prediction. The project is in alpha, and single-cell RNA-seq processing is listed as planned rather than available.

The server is intended for computational analyses that produce concise, agent-readable results alongside files on disk. A request can return summaries, messages, output paths, suggestions, and—in the case of failures—a hint and retryability flag. The same operations are available through the BioHarbor command-line interface, so an analysis can be run with or without an agent.

## How it works

Each tool call is recorded in SQLite with its parameters, software versions, timings, and the GPU used when applicable. BioHarbor also writes a `provenance.json` file next to generated outputs. Before work is queued, it checks the request, required binaries, and configured databases so invalid jobs fail before waiting for compute.

Short operations run inline. Longer work returns a `job_id` instead of holding the client connection until completion. Agents can inspect, list, or cancel jobs, and can retrieve files after a run. GPU placement checks free memory and utilization through NVML or `nvidia-smi`, retains headroom, and tracks reservations for jobs it has started.

## Setup and configuration

Install the package with `pip install bioharbor`. Run `bioharbor doctor` to check the Python environment, GPUs, workspace, and external tools. Homology searches require a local reference database and MMseqs2; the README shows `bioharbor setup-db swissprot` for setting up Swiss-Prot. ESMFold support is installed with the optional `bioharbor[esmfold]` extra, and RTX 50xx hardware requires CUDA 12.8 or newer PyTorch according to the project documentation.

Use `bioharbor serve` for a standard stdio MCP server. Streamable HTTP is available with `bioharbor serve --http`; for a remote GPU host, the documented approach is to bind locally on the server and connect through an SSH tunnel. Client setup commands are provided for Claude Desktop, Cursor, Claude Code, and Codex. The server can also be added to Biomni using its MCP configuration interface.

## Tools and capabilities

Core analysis tools include:

- `seq_stats` for validation, sequence type, length, GC percentage, and molecular weight.
- `translate_sequence` for DNA or RNA translation in one or all six frames.
- `find_orfs` for ORF discovery on both strands with coordinates.
- `search_homologs` for MMseqs2 searches against local databases using proteins or translated DNA.
- `predict_structure` for ESMFold models, pLDDT bands, low-confidence regions, and pTM values.

Runtime tools include job inspection and cancellation, tool and database descriptions, GPU status, and file reading.

## Limitations and notes

BioHarbor MCP server is alpha software at version 0.1. HTTP mode currently has no authentication, so the README advises keeping it on `127.0.0.1` and using an SSH tunnel for remote access. Structure prediction depends on suitable GPU software and hardware, while homology search depends on MMseqs2 and a prepared local database. The single-cell RNA-seq pipeline is planned and should not be treated as an available tool.

_Full upstream README: https://allmcps.com/mcp/bioharbor/readme_