LLM Integration (MCP)¶

atb includes a built-in Model Context Protocol server, allowing LLMs to query the AllTheBacteria database directly through natural language.

Two transport modes: - stdio (default) - for Claude Code, Claude Desktop, Cursor, VS Code Copilot, Windsurf, OpenAI Codex CLI - HTTP/SSE (--http :8080) - for ChatGPT, OpenAI Responses API, remote clients

Tools exposed:

Tool	Description
`atb_query`	Search genomes by species, genus, quality, N50
`atb_amr`	Query AMR resistance genes by species and drug class
`atb_mlst`	Query MLST scheme, ST, and allele calls
`atb_info`	Get full metadata for a specific sample
`atb_stats`	Database summary statistics
`atb_species_list`	List available species with genome counts

Quick try (no install needed, requires Go)¶

# Claude Code - runs directly from GitHub, available in all projects
claude mcp add --scope user atb -- go run github.com/allthebacteria/atb-cli/cmd/atb@latest mcp

First call takes ~10s to compile; cached after that.

Setup¶

# 1. Install atb
curl -fsSL https://raw.githubusercontent.com/allthebacteria/atb-cli/main/install.sh | bash

# 2. Fetch the database and build the index
atb fetch

See Fetching & indexing data for details on data directory options and selective table downloads.

stdio mode (Claude, Cursor, Codex CLI)¶

Claude CodeClaude DesktopCursorOpenAI Codex CLI

# Claude Code (available globally in all projects)
claude mcp add --scope user atb -- atb mcp

# Claude Code (current project only)
claude mcp add atb -- atb mcp

{
  "mcpServers": {
    "atb": {
      "command": "atb",
      "args": ["mcp"]
    }
  }
}

Add to ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows.

# Cursor (Settings > MCP Servers > Add)
# Command: atb mcp

# OpenAI Codex CLI (~/.codex/config.toml)
[mcp_servers.atb]
command = "atb"
args = ["mcp"]

Restart required

After adding, restart your client for the MCP server to become available.

HTTP/SSE mode (ChatGPT, OpenAI API, remote clients)¶

# Start the HTTP/SSE server
atb mcp --http :8080

Then configure your client with the SSE endpoint URL: - ChatGPT: Settings > Connected apps > Add MCP server > http://your-host:8080/sse - OpenAI Responses API: Use server_url: "http://your-host:8080/sse" in the MCP tool config

For public access, deploy with Docker or use a tunnel:

# Docker (auto-downloads data on first run)
docker compose up -d
# SSE endpoint: http://localhost:8080/sse

# Or quick public URL with ngrok
atb mcp --http :8080 &
ngrok http 8080

Non-default data location

If your data is in a non-default location, add --data-dir /your/path to all commands above.

What you can ask your LLM¶

Once connected, you can ask natural language questions like:

"How many Salmonella genomes are in the database?"
"Find me 20 high-quality E. coli genomes with N50 > 200000"
"What beta-lactam resistance genes does Klebsiella pneumoniae have?"
"Show me all metadata for sample SAMD00000355"
"What are the top 10 species by genome count?"

The LLM will call the appropriate atb tools and interpret the results for you.

For the full atb mcp command reference, see CLI: atb mcp.