MCP server
For any MCP client: Claude, Cursor, Codex, LangChain, CrewAI, smolagents, n8n. No code to write.
SDK tools
doconda.tools(): tools ready for function calling with Claude, OpenAI, Gemini, Mistral, Bedrock, Vertex AI or the AI SDK.REST API
For everything else: an HTTP request with your API key.
The tools
The MCP server and the SDK (doconda.tools() in TypeScript and Python) give your agent the same tools:
create_document and edit_document wait until the document is finished, then return a short summary the agent can
pass on to the user:
{ "error": "<code>", "message": "…" }. On a validation error, fields names each wrong argument. The codes are the
API’s error codes.
By stack
Key facts
- Billing — Everything the agent does counts as if you did it with the API: same project, same region, same price.
- Waiting —
create_documentandedit_documentwait for the finished document, up to about 4 minutes: about a minute atquality: "fast", several minutes atstandardorbest. A document that takes longer comes back with itsid,statusand anote; the agent callsget_documentlater for the result and its links. MCP clients need a per-call timeout of 300 s. Claude’s connector gives up on a call after about a minute, so the hosted MCP server answers it after about 50 s. With the SDK tools,waitSeconds(wait_secondsin Python) sets the wait. - Links — File links expire after 5 minutes;
get_documentreturns fresh ones. An artifact’s page works until the document expires. - Files — Uploads are up to 20 MB. Only the local MCP server, or
doconda.tools({ localFiles: true })(tools(local_files=True)in Python), reads files from disk; otherwise the agent passes a publicurlorcontent_base64. - Conversions — The first 1,000
convert_filecalls each month are free, then €0.01 each.
Next steps
Connect an MCP client
Local or remote, with an API key or by signing in.
Use the SDK tools
doconda.tools() in TypeScript and Python.Read the API reference
Every endpoint the tools call.