Quickstart

Two paths. The MCP one is what most agents use; the HTTP one proves the API is reachable without any client at all.

Talk to the public API first

Nothing here needs a credential, so it is a clean way to confirm you can reach Mitosis:

curl -s https://mitosislabs.ai/api/v1 | jq

The index self-identifies its environment, which matters once you start pointing the same code at dev.mitosislabs.ai:

{ "environment": "production", "endpoints": { "...": "..." } }

Connect an MCP client

The product MCP server is Streamable HTTP at https://mitosislabs.ai/api/mcp. It supports OAuth 2.1 with dynamic client registration, so a client that speaks MCP auth needs no pre-provisioned key. The user approves access in a browser.

claude mcp add --transport http mitosis https://mitosislabs.ai/api/mcp

{
  "mcpServers": {
    "mitosis": {
      "url": "https://mitosislabs.ai/api/mcp"
    }
  }
}

npx -y -p @mitosislabs/sdk@latest mi login

The first tool call that touches a memory triggers the OAuth flow. Public tools (get_pricing, get_platform_status, search_docs, list_skills) answer before you sign in at all.

Ask the graph a question

Once connected, the memory tools appear. The one to reach for first is cortex_ask. It answers in natural language and cites the nodes it used:

cortex_ask({ question: "What did we decide about pricing last quarter?" })

Every result carries citations and a cited_graph_url that deep-links to those exact nodes in the user's own graph.

Then write something back

Memory is not read-only. An agent that learns a durable fact should store it:

cortex_remember({ text: "Prefers async design review over live meetings." })

That fact is embedded and linked into the same graph, so a later cortex_ask from any client can retrieve it.

Where to go next