FunctionTool, or connect the MCP server.
Capabilities
- Create a Word, PowerPoint, Excel or PDF file, or an artifact, from a request or from Markdown the agent wrote.
- Edit a file: only what is asked changes, and
get_reportlists every change. - Convert a file to another format with no AI, for example any file to Markdown so the agent can read it.
- Read what people saved in an artifact, such as poll votes.
Setup
This integration uses a Doconda API key inDOCONDA_API_KEY (from API keys). The Agents SDK
reads your OpenAI key from OPENAI_API_KEY.
With the Python SDK
doconda.tools() is JSON Schema, which the Agents SDK takes in FunctionTool:
result.final_output is the agent’s answer, with the memo’s download links. Set strict_json_schema=False: strict
mode makes every field required, and most of Doconda’s are optional (prompt or markdown, style, name…).
asyncio.to_thread keeps the event loop free while a document is made.
With MCP
client_session_timeout_seconds=300 raises the wait per call from the default 5 s. A tool call returns within about 4
minutes: with the finished document, or with its id and status when it takes longer, for the agent to check with
get_document. To run the server locally instead of the remote one:
Tools
The agent gets the eight Doconda tools; see the tools for what each one does.doconda.tools() returns them with name, description, input_schema (JSON Schema) and run. A failed call returns
{ "error": "<code>", "message": "…" } instead of raising, so the model can correct itself.
Troubleshooting
- The MCP call times out — The default 5 s per call is shorter than a document takes. Set
client_session_timeout_seconds=300. Doconda: pass api_key or set DOCONDA_API_KEY.— The SDK found no key: setDOCONDA_API_KEY, or passapi_keytoDoconda.- The download link stopped working — File links expire after 5 minutes; the agent calls
get_documentfor fresh ones.
Useful links
- Python SDK.
- MCP server: local and remote setup.
- doconda-examples on GitHub: ready-to-run scripts.
- API reference.