Guides

For agents

The public API is offered to agents in two complementary forms: an MCP server for acting, and machine-readable docs for reading first.

The MCP server

The platform mounts an MCP server at:

Code
https://YOUR_PLATFORM_ORIGIN/api/v1/mcp

Transport: streamable HTTP.
Auth: the same bearer header as the REST API — the caller's Authorization: Bearer atk_... is forwarded onto the underlying HTTP call, so a tool call runs with exactly the reach of that key: the same organization, the same flags. A call with no header goes out unauthenticated and is refused.
Toolset: exactly the operations in this reference. The tools are derived from the same public slice of the API that this site renders — there is no separate "agent API", and an agent can never do more through a tool than the key could do over HTTP.

Tool names are the operations' operationIds (for example list_agents, create_agent), the same identifiers shown in the API Reference.

Machine-readable docs

This site publishes its content for models alongside the HTML:

/docs/how-it-fits.md — read this first: it is the page that says which operations a job needs, and in which order. The recipes under /docs/recipes/ are the sequences themselves, each run as a test.
/docs/tools-and-parameters.md — read this before creating or editing a tool instance: what a built-in accepts comes from list_builtins and never from an existing row, and a fixed_parameters edit replaces the whole set.
/docs/openapi.public.json — the exact OpenAPI document this reference renders: every public operation and schema, and nothing else. This is the file to read for the operations.
/docs/llms.txt — an index of the guide pages, with links.
/docs/llms-full.txt — the guides' full content in one file.
Each guide page also has a markdown twin at the same path with an .md suffix.

The published document and the MCP toolset are the same derivation of the API, so what an agent reads here and what it can call there cannot disagree about what exists.

Last modified on November 4, 2026

Authentication Errors

The MCP server

The platform mounts an MCP server at:

Code

https://YOUR_PLATFORM_ORIGIN/api/v1/mcp

Transport: streamable HTTP.

Auth: the same bearer header as the REST API — the caller's Authorization: Bearer atk_... is forwarded onto the underlying HTTP call, so a tool call runs with exactly the reach of that key: the same organization, the same flags. A call with no header goes out unauthenticated and is refused.

Toolset: exactly the operations in this reference. The tools are derived from the same public slice of the API that this site renders — there is no separate "agent API", and an agent can never do more through a tool than the key could do over HTTP.

Tool names are the operations' operationIds (for example list_agents, create_agent), the same identifiers shown in the API Reference.

Machine-readable docs

This site publishes its content for models alongside the HTML:

/docs/how-it-fits.md — read this first: it is the page that says which operations a job needs, and in which order. The recipes under /docs/recipes/ are the sequences themselves, each run as a test.

/docs/tools-and-parameters.md — read this before creating or editing a tool instance: what a built-in accepts comes from list_builtins and never from an existing row, and a fixed_parameters edit replaces the whole set.

/docs/openapi.public.json — the exact OpenAPI document this reference renders: every public operation and schema, and nothing else. This is the file to read for the operations.

/docs/llms.txt — an index of the guide pages, with links.

/docs/llms-full.txt — the guides' full content in one file.

Each guide page also has a markdown twin at the same path with an .md suffix.

The published document and the MCP toolset are the same derivation of the API, so what an agent reads here and what it can call there cannot disagree about what exists.

Last modified on November 4, 2026