Connect any AI client to BioMate

What are connectors?

BioMate connectors let AI assistants call BioMate's scientific-workflow platform on your behalf. When you ask Claude Code to "run RNA-seq on these FASTQs," it calls the run_workflow tool rather than generating code — the run executes on BioMate cloud, and live phase and step updates stream back into the chat.

The connector stack has two transport families:

• MCP stdio — Claude Code, Claude Desktop, Cursor, and Codex CLI use a local proxy process. The AI client speaks JSON-RPC over stdio to a BioMate-issued binary that relays calls to dev-public.biomate.ai.
• REST/webhook — ChatGPT, Slack, and Feishu call BioMate's REST API at dev-public.biomate.ai directly over HTTPS. No local installation needed.

All surfaces authenticate with the same OAuth 2.1 + PKCE flow against dev-public.biomate.ai/oauth. Your BioMate account, runs history, and workflow catalog are shared across every surface.

Supported surfaces

Available tools (19)

All surfaces expose the same tool set, sourced from a single canonical manifest (tools_manifest.py). Tools are annotated with read-only and destructive hints so the AI client can present appropriate confirmation prompts.

Streaming agentic tool

Workflow primitives

Data and knowledge

Outputs, analysis, reporting

End-to-end use cases

Each example below shows the full information flow — from a natural-language prompt to a delivered result — across a specific connector surface. Every call is a real API call against dev-public.biomate.ai.

RNA-seq differential expression — Claude Code

A researcher asks Claude Code to run a differential expression analysis on uploaded FASTQ files stored in S3.

# User types in Claude Code terminal:
Run RNA-seq differential expression on
s3://biomate-demo/treated_R1.fq.gz vs s3://biomate-demo/control_R1.fq.gz.
Use GRCh38, DESeq2. Return a volcano plot and the top 50 DEGs.

Information flow:

1. Search — Claude Code calls search_workflow("RNA-seq differential expression DESeq2") → BioMate returns ranked matches; top hit is nf-core/rnaseq with DESeq2 post-processing.
2. Spec — Claude calls get_workflow_spec("nf-core/rnaseq") → receives required params (input, genome, contrasts), optional params, default QC thresholds, and cost estimate (~$0.40 on AWS Batch).
3. Run — Claude calls run_workflow({ workflow: "nf-core/rnaseq", params: { input: "s3://biomate-demo/…", genome: "GRCh38", contrasts: "treated_vs_control" }, stream: true }) → BioMate queues the job on AWS Batch and returns a run_id.
4. Stream — watch_run(run_id) emits notifications/progress events: FASTQC → STAR alignment → featureCounts → DESeq2 → MultiQC. Claude renders each phase inline as it completes.
5. Results — get_run(run_id) returns output file list (volcano.png, degs.csv, multiqc_report.html). Claude calls preview_file("s3://…/volcano.png") and displays the plot inline.
6. Report — export_report(run_id, format="pdf") → signed S3 URL to a methods PDF with QC metrics and DEG table, ready for the lab notebook.

ADMET screening with auto-loop QC — Claude Code

A medicinal chemist screens a compound library and lets BioMate's auto-loop correct hERG-failing structures.

# User types:
Screen these SMILES for hERG, CYP3A4, and oral bioavailability.
Retry any hERG-failing compounds with reduced lipophilicity.

CC1=CC=C(C=C1)S(=O)(=O)N   aspirin-like
CN1C=NC2=C1C(=O)N(C(=O)N2C)C   caffeine

Information flow:

1. Search — search_workflow("ADMET hERG CYP3A4 oral bioavailability") → returns admet_prediction_suite.
2. Run — run_workflow({ workflow: "admet_prediction_suite", params: { smiles_list: ["CC1=CC=C…", "CN1C=NC2…"], endpoints: ["hERG","CYP3A4","Foral"] } }).
3. QC gate fires — Compound 2 exceeds hERG IC₅₀ threshold (0.41 µM vs 1.0 µM limit). BioMate auto-loop triggers: explain_error(run_id, gate="hERG") returns root-cause + suggested param change (reduce cLogP by 1.2).
4. Auto-retry — BioMate resubmits with modified SMILES (desethyl analog). Claude shows a diff card: original → revised.
5. Results — get_run(run_id) → all endpoints pass. export_report(run_id, format="pdf") → CRO-ready ADMET dossier.

GEO accession → scRNA-seq — Slack

A bioinformatics team member pastes a GEO accession in Slack and BioMate runs the full pipeline automatically.

@BioMate Analyze GSE142256 — MG patient PBMCs, 10x Chromium.
Cluster by cell type and find differentially expressed genes in T cells.

Information flow:

1. Slack event — Slack posts the app_mention event to the BioMate bot endpoint at dev-public.biomate.ai/integrations/slack/command. The bot verifies the Slack signing secret and responds 200 OK immediately (async processing).
2. Accession resolve — Bot calls resolve_accession("GSE142256") → BioMate identifies this as a 10x scRNA-seq dataset, returns pre-filled workflow params: nf-core/scrnaseq + input: "s3://biomate-staging/geo/GSE142256/", genome: GRCh38.
3. Data stage — fetch_public_data("GSE142256") pulls raw FASTQ files from GEO/SRA into the BioMate S3 workspace. Bot posts "Staging data…" card update in the Slack thread.
4. Run — run_workflow({ workflow: "nf-core/scrnaseq", … }) → queued on AWS Batch. Bot posts "Run started" card with run ID and estimated completion time.
5. Completion — On completion, bot calls get_run(run_id) and analyze_results(run_id, question="Top markers per cluster, DEGs in T cells"). Bot posts a summary card with cluster UMAP, top markers table, and an Open in BioMate button deep-linking into the run panel.

Literature + database query — ChatGPT

A scientist uses the BioMate Custom GPT to look up a gene target, search the literature, then run a pathway analysis — all in one conversation.

Tell me everything about CD40LG as an autoimmune drug target.
Then run a pathway enrichment on the top 50 CD40L-correlated genes from GTEx.

Information flow:

1. Database lookup — ChatGPT calls query_database({ db: "UniProt", query: "CD40LG human" }) → returns UniProt entry P29965: gene aliases, function, 12 disease associations, 3 known inhibitors.
2. Literature search — search_literature({ query: "CD40LG autoimmune disease target 2020-2025", sources: ["PubMed","Semantic Scholar"] }) → returns 8 high-relevance papers with abstracts; ChatGPT summarizes mechanism and clinical evidence.
3. Data fetch — fetch_public_data("GTEx v10 CD40LG expression") → stages the GTEx expression matrix into BioMate workspace.
4. Workflow search — search_workflow("pathway enrichment clusterProfiler") → returns clusterProfiler_enrichGO workflow.
5. Run — run_workflow({ workflow: "clusterProfiler_enrichGO", params: { gene_list: top50, organism: "human", ont: "BP" } }) → AWS Batch job.
6. Report — export_report(run_id, format="markdown") → ChatGPT embeds the top GO terms directly into its reply with a download link to the full PDF.

PBPK modeling for IND submission — Claude Desktop

A pharmacokineticist builds a full IND evidence package: ADMET → PBPK → regulatory narrative, all from Claude Desktop.

# User in Claude Desktop:
Run ADMET on our lead compound (SMILES: CC(C)Cc1ccc(cc1)C(C)C(=O)O),
then PBPK in a 70 kg adult at 100 mg oral dose,
then generate a CTD §2.6.1 PK narrative for IND filing.

Information flow:

1. ADMET run — run_workflow("admet_prediction_suite", { smiles: "CC(C)Cc1ccc…" }) → returns Foral 68%, hERG IC₅₀ 8.4 µM (pass), CYP3A4 substrate (yes). Results stored as admet_run_id.
2. PBPK run — run_workflow("pbpk_11compartment", { smiles: "…", dose_mg: 100, route: "oral", bw_kg: 70, fu: 0.23 }) → 11-compartment ODE model; Cmax 3.2 µg/mL at 1.4 h, AUC₀–₂₄ 18.4 µg·h/mL. Results as pbpk_run_id.
3. Memory — recall_memory({ context: "PBPK similar compounds" }) → retrieves prior runs for structurally similar compounds to cite precedent in the narrative.
4. Narrative generation — run_workflow("ctd_261_narrative_generator", { admet_run: admet_run_id, pbpk_run: pbpk_run_id }) → LLM drafts the §2.6.1 PK summary using actual computed values, formats citations, writes Methods and Results subsections.
5. Export — export_report(narrative_run_id, format="pdf") → IND-ready DOCX/PDF with SHA-256 audit manifest (claims_index.json) for traceability. Signed S3 URL returned; Claude Desktop renders a download link inline.

Claude Code

The Claude Code connector runs as a local MCP server (stdio transport). It installs in 30 seconds and writes its config into ~/.mcp.json.

Install via CLI

npx @biomate/connect claude-code

The installer opens your browser, completes the OAuth 2.1 + PKCE flow against dev-public.biomate.ai, and writes the server entry into your Claude Code config. Restart Claude Code to activate.

Try it

# ADMET screening
Screen aspirin and caffeine for hERG inhibition and CYP3A4 metabolism.

# RNA-seq differential expression
Run RNA-seq differential expression on s3://biomate-demo/treated_R1.fq.gz
and s3://biomate-demo/control_R1.fq.gz. Compare treated vs control, GRCh38.

# Cryo-EM refinement
Refine s3://biomate-demo/particles.cs with CryoSPARC homogeneous refinement.
Apply C2 symmetry.

Claude Desktop

Claude Desktop supports MCP stdio servers via the claude_desktop_config.json file.

Install via CLI

npx @biomate/connect claude-desktop

The installer handles OAuth and writes the server entry into ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). Restart Claude Desktop to activate.

Manual config

If you prefer to configure manually, add this entry under mcpServers in claude_desktop_config.json:

{
  "mcpServers": {
    "biomate": {
      "command": "npx",
      "args": ["-y", "@biomate/mcp-server"],
      "env": {
        "BIOMATE_API_KEY": "<your-api-key>"
      }
    }
  }
}

ChatGPT

BioMate integrates with ChatGPT as a custom GPT action. The /tools/* adapter routes run on dev-public.biomate.ai — no local installation needed.

Slack

The BioMate Slack integration adds a bot to your workspace. Mention @BioMate in any channel to start a workflow, check run status, or query the database.

Example interactions

@BioMate Run GATK variant calling on s3://my-bucket/sample.bam, hg38

@BioMate What's the status of my last run?

@BioMate Look up UniProt P04637 and summarize cancer mutations

Feishu / Lark

The BioMate Feishu integration connects as a Feishu open-platform bot. It supports interactive message cards that display run status and findings inline.

Authentication

All surfaces use OAuth 2.1 + PKCE. No shared secrets or passwords are stored by the connector — only short-lived access tokens (30 min) and hashed refresh tokens.

Scopes

Revoke any surface's access at any time at biomate.ai/account/connectors. Revoking one surface does not affect others.

API keys

For server-to-server use (no browser), generate a long-lived API key at biomate.ai/account/api-keys. Pass it as Authorization: Bearer <key> or set it in the BIOMATE_API_KEY environment variable.

Manual configuration

If npx @biomate/connect is unavailable, set the server entry by hand. For Claude Code, add this to ~/.mcp.json:

{
  "mcpServers": {
    "biomate": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@biomate/mcp-server"],
      "env": {
        "BIOMATE_API_KEY": "bm_sk_..."
      }
    }
  }
}

The BIOMATE_API_KEY value can be found at biomate.ai/account/api-keys.

FAQ

Does the connector send my data to Anthropic?

No. The connector calls BioMate's API directly. Anthropic (Claude Code / Claude Desktop) only sees the tool names, their inputs, and their outputs — the same data visible in your chat session. See the privacy policy for full details.

Can I use the connector with a self-hosted BioMate instance?

Yes. Set BIOMATE_API_URL=https://your-instance.example.com in the environment before running npx @biomate/connect.

Is the connector open source?

Yes. The server code and all surface adapters are at github.com/bioMate-AI/biomate-connector under the MIT license.

What happens if a run exceeds my quota?

The run_workflow tool returns an error with the quota details. The AI client surfaces this as a readable message. Upgrade your plan at biomate.ai/billing.

Can multiple team members share the same BioMate connector?

Each team member authenticates with their own BioMate account. Runs, files, and project membership are governed by BioMate's project-level access controls — not the connector.