MCP server

The ivybloom MCP server gives an AI agent the platform's tools over the Model Context Protocol. An agent in Claude Code, Claude Desktop, or your own MCP client can submit jobs, search compounds, manage sessions, and generate passports.

It runs on the same backend and job history as Chat and the ivybloom CLI, so agent work shows up next to everything else. The CLI's MCP guide has the full tool catalog and client configs.

What MCP is

The Model Context Protocol is an open standard for connecting AI agents to external tools. An MCP server advertises a set of tools an agent can call; the agent picks tools, sends arguments, and reads results back — all over a structured JSON-RPC channel.

Connecting

The MCP server is part of the ivybloom CLI. Once the CLI is installed and authenticated, start the server over stdio:

# authenticate once
ivybloom auth login --browser

# start the MCP server over stdio (core toolset)
ivybloom mcp serve

# expose every tool domain
ivybloom mcp serve --toolset all

The server speaks JSON-RPC over standard input/output and is meant to be launched by an agent or IDE rather than used interactively. To wire it into Claude Desktop, add it to claude_desktop_config.json:

{
  "mcpServers": {
    "ivybloom": {
      "command": "ivybloom",
      "args": ["mcp", "serve"]
    }
  }
}

The same command / args shape works in a Claude Code project .mcp.json file, or register it with claude mcp add ivybloom -- ivybloom mcp serve. Any other MCP-capable client takes the same command. Add "--toolset", "all" to args for the full catalog. See the CLI MCP guide for Python client examples and the full integration walkthrough.

What is exposed

Tools are named domain_action (for example jobs_submit, compounds_search, passport_generate). By default the server advertises the core toolset: the jobs, tools, projects, results, storage, stats, suggest and orchestrate domains. --toolset research, --toolset admin and --toolset all widen it, and --domains jobs,compounds picks exact domains. Advertising every tool costs an agent roughly 20k tokens of context, which is why core is the default.

The catalog differs between CLI versions. Run ivybloom mcp tools --toolset all to list what your installed binary advertises. The main families:

Agents can also read URI-addressable resources (jobs://{id}, projects://list, sessions://list and others; ivybloom mcp info lists them) to inspect data without executing a tool, and two guided prompt templates — analyze_protein and drug_screening — walk an agent through common workflows.

Shared backend and auth

The MCP server is not a separate system. It forwards to the same backend as Chat and the CLI, so it shares one job history, one credit balance, and one set of credentials. Authenticate the CLI once (ivybloom auth login) and the server inherits that identity; jobs an agent submits over MCP are the same jobs you see in the web app and the CLI.

Because long-running compute (folding, docking) can exceed an agent's request timeout, tools like jobs_submit return a job ID immediately — the agent should poll jobs_status for completion rather than waiting synchronously.

Access tier

The MCP server is not gated by plan. It needs the ivybloom CLI, installed and signed in, on any account, including the free Seed plan; there is no anonymous MCP path to the platform tools. The REST API needs only an API key from a free account, and the keyless first call is the guest curl request. Your plan decides the credits, model families and compute tools behind each call; see Plans & tiers.

A separate, read-only docs server runs at https://www.ivybiosciences.com/api/mcp/docs (JSON-RPC over HTTP, no sign-in). It offers two tools, search_docs and get_doc_markdown, so an agent can read these docs without installing the CLI.

Read next