MCP server

Cruq AI exposes a Model Context Protocol server so agents and MCP clients (Claude, Cursor, and others) can drive the platform directly, using the same tools as the REST API.

Endpoint

URL: https://api.cruq.ai/mcp
Transport: Streamable HTTP
Auth: the same API key as the REST API, sent as a bearer token: Authorization: Bearer cruq_sk_.... A runtime key gets the tools that run, read and search. A management key (cruq_mk_...) also gets the build tools, which create and change agents.

The server is stateless, so any MCP client that speaks Streamable HTTP can connect without session setup.

Connecting a client

Point your MCP client at the endpoint with an Authorization header. For a client configured via JSON:

{
  "mcpServers": {
    "cruq": {
      "url": "https://api.cruq.ai/mcp",
      "headers": { "Authorization": "Bearer cruq_sk_..." }
    }
  }
}

Tools

Tool	Description
`list_agents`	List the agents in your workspace.
`run_agent`	Run an agent’s published version and return its final text and `thread_id`. Pass the `thread_id` back to continue the conversation.
`list_runs`	List recent runs.
`get_run`	Get a run and its tool-call step trace.
`search_memory`	Semantically recall stored memories.
`add_memory`	Store a memory for later recall.
`search_knowledge`	Search knowledge bases (RAG).
`list_knowledge_bases`	List knowledge bases.
`list_tools`	List the workspace tool catalog.

Each tool maps to the same operation as the REST API; see the OpenAPI spec for the full schema.

Build tools

With a management key, the server also lists tools for building agents from a conversation. This is what the Agent Builder team runs on, and any MCP client can use them the same way.

Tool	Description
`list_capabilities`	What agents here can use: connected model providers, builtin tools, tools and connected integrations, integrations not yet connected, existing agents and knowledge bases.
`get_agent`	An agent’s current model, instructions, builtins, tools, subagents and starter prompts.
`create_agent`	Create an agent from plain fields and publish it. Fails if the name is taken.
`update_agent`	Change an agent and publish a new version. Omitted fields keep their value.
`create_knowledge_base`	Create a knowledge base, or return the one with that name.
`add_document`	Add a text document to a knowledge base. It is searchable once the background ingest finishes.

create_agent and update_agent take tools by name, not by id:

{
  "name": "Order Desk",
  "description": "Answers order questions for Meridian Foods",
  "instructions": "You answer order questions. Ask for the email on the order first.",
  "model": "openrouter/openai/gpt-4o-mini",
  "builtins": ["knowledge_search", "ask_human"],
  "tools": ["Tavily"],
  "subagents": ["Billing Agent"],
  "sample_prompts": ["Where is order 1042?"]
}

builtins are builtin tool keys. browser also needs browser_domains, the sites it may open.
tools are tool names as list_capabilities shows them. A connector’s own name ("Tavily") picks all of its tools.
On update_agent, a builtins, tools or subagents list replaces that list only, so adding a builtin keeps the agent’s integrations.
Results carry warnings, for example a model whose provider is not connected, or an integration that is attached but not connected yet.

None of the build tools delete anything. Agents with routers or several agent nodes can be read with get_agent but must be edited in the dashboard.

Getting Started

Demos

Concepts

Reference

Endpoint

Connecting a client

Tools

Build tools

​Endpoint

​Connecting a client

​Tools

​Build tools

Endpoint

Connecting a client

Tools

Build tools