BioHarbor
BioHarbor
Run real bioinformatics from your AI agent. Reliable, reproducible, on your own GPUs.
π§ͺ Alpha (v0.1). Sequence tools, homology search (MMseqs2) and structure prediction (ESMFold, validated on RTX 5090) work. Feedback welcome β see the roadmap.
BioHarbor is an MCP server that lets AI agents such as Claude, Cursor and Codex execute bioinformatics tools β 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.

Why another bio MCP server?
Most bio MCP servers wrap databases (UniProt, PDB, PubMedβ¦). Use them β BioHarbor complements them by running the compute:
Database MCP servers | BioHarbor | |
Runs real analyses (search, fold, cluster) | β | β |
Validates inputs before burning GPU time | β | β |
GPU-aware queue, polite on shared GPUs | β | β |
Long jobs return a | β | β |
Compact summaries + files on disk (saves tokens) | β | β |
Provenance for every run, export to a pipeline | β | β (export: planned) |
Related MCP server: BioLab MCP Server
Quick start
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 CodexFor structure prediction on a GPU: pip install "bioharbor[esmfold]" β 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 Desktop |
|
Cursor |
|
Codex (CLI, IDE extension, app) |
|
Biomni (Stanford's biomedical agent) |
|
Anything else | run |
Long-running tools return a job_id within ~20 s instead of blocking, so they stay
within every client's tool-call timeout.
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:
bioharbor tools list
bioharbor run find_orfs sequence=@contig.fa min_aa=100 --brief # human-readable
bioharbor run seq_stats sequence=MKTAYIAKQRQISFVKSHFSRQ
bioharbor jobsShared 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:
# 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-serverβ οΈ HTTP mode has no authentication yet (on the roadmap), so keep it on
127.0.0.1and use the tunnel. Details: docs/connect-clients.md.
Tools
Tool | What it does | Runs |
| Validate sequences; type, length, GC%, molecular weight | inline |
| DNA/RNA β protein, one or all six frames | inline |
| Longest ORFs on both strands, with coordinates | inline |
| MMseqs2 search (protein, or translated DNA) vs local DBs | job |
| ESMFold structure, pLDDT bands, low-confidence regions, pTM | job (GPU) |
| scanpy QC β clustering β markers | planned |
Runtime tools: get_job, list_jobs, cancel_job, describe_tool, list_databases,
gpu_status, read_file.
How it works
Agent ββMCPβββΆ validate input ββΆ inline? ββyesβββΆ run ββ
β no βββΆ provenance + summary ββΆ Agent
βΌ β
job queue (SQLite) ββΆ GPU placement β
(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.jsonnext 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 ahintand aretryableflag.
Details: docs/design.md.
Writing a tool
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
Available Tools
12 toolscancel_jobB
Cancel a job that is still queued or waiting for a GPU.
Inputs: job_id (required). Returns: object
describe_toolC
Full input schema, version and resource needs of a BioHarbor tool.
Inputs: name (required). Returns: object
find_orfsB
Find open reading frames (ATG to stop) on both strands.
Returns the longest ORFs in DNA/RNA with coordinates (1-based, on the forward strand) and writes all ORFs to a FASTA file.
Inputs: sequence (required), top_n, min_aa, both_strands. Returns: object
get_jobB
Status and (when finished) result of a background job.
Inputs: job_id (required), wait_seconds. Returns: object
gpu_statusA
Live GPU memory and utilisation (includes other users' processes).
Inputs: no named parameters. Returns: object
list_databasesB
Sequence databases installed for search_homologs, and ones that can be added.
Inputs: no named parameters. Returns: object
list_jobsA
Recent jobs, newest first (results omitted; use get_job for details).
Inputs: limit, status. Returns: object
predict_structureA
Predict 3D protein structure with ESMFold on a GPU.
Returns per-protein mean pLDDT, confidence bands, low-confidence regions and pTM; PDB files are written to disk (B-factor column = pLDDT).
Runs as a background job; may return a job_id to poll with get_job.
Inputs: sequence (required), num_recycles. Returns: object
read_fileB
Read (the start of) a text output file produced by a BioHarbor tool.
Inputs: path (required), max_bytes. Returns: object
search_homologsA
Find homologous proteins with MMseqs2 (Swiss-Prot, PDB, ...).
Searches a local database (default Swiss-Prot). Returns the best hits per query with identity, E-value, coverage, organism and description; the full hit table is written to a TSV file.
Runs as a background job; may return a job_id to poll with get_job.
Inputs: sequence (required), top_n, evalue, database, +3 more. Returns: object
seq_statsA
Validate sequences: type, length, GC%, molecular weight.
Reports dna/rna/protein type, length, GC content (nucleotides) or molecular weight (proteins) and composition. A cheap first step before heavier tools.
Inputs: sequence (required). Returns: object
translate_sequenceA
Translate DNA/RNA to protein, in one frame or all six.
Uses the standard genetic code; stop codons appear as '*'.
Inputs: sequence (required), frame, to_stop. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelogβ¦
TDQS
Scored across 12 tools
Each tool targets a distinct operation: sequence statistics, translation, ORF finding, homology search, structure prediction, and job management are clearly separated. No overlapping purposes; get_job, list_jobs, and cancel_job are distinct job operations.
Names uniformly use snake_case and mostly follow a verb_noun pattern (e.g., translate_sequence, find_orfs, get_job). Minor deviation: seq_stats and gpu_status are noun phrases, but overall consistent and readable.
12 tools is well within the ideal 3-15 range, each tool serves a clear purpose in the bioinformatics workflow without redundancy. The set is well-scoped for the domain.
Covers a sequence analysis pipeline from validation to structure prediction, with job management and resource introspection. Minor gap: no explicit tool for retrieving or loading input sequences (e.g., from files or databases), though read_file exists for outputs.
Maintenance
Related MCP Connectors
Protein analysis: ESM-2/ESMC embeddings, mutation scoring, landscape scans, ESMFold structure.
Find open biology and chemistry AI models, bioinformatics tools and databases, and run them online.
Turn any LLM into your lab assistant: search samples, track experiments, analyze data with AI.
Fold proteins, co-fold ligands with binding affinity, design binders and embed sequences.
Related MCP Servers
- FlicenseNot gradedqualityCmaintenanceEnables AI-powered protein design, including binders, peptides, and custom proteins, with GPU-accelerated Docker deployment and async job management.2-
- AlicenseNot gradedqualityCmaintenanceUnified MCP server providing AI-agent-ready access to AlphaFold, PubMed, ChEMBL, Ensembl, and 37+ scientific databases.MIT
- AlicenseNot gradedqualityCmaintenanceMCP-native scientific skills for reproducible computational biology and AI-driven drug-discovery workflows. It combines deterministic scientific tools with an MCP server to give AI agents real computational capabilities.Apache 2.0
- AlicenseNot gradedqualityDmaintenanceEnables LLM agents to design proteins, predict structures, score interfaces, and run molecular dynamics simulations through a unified interface.1Apache 2.0