The tool list is not fixed by a vendor — it is whichever UDFs you left visible on a canvas, so the agent's reach is edited the same way you edit the pipeline. It also keeps the heavy part in place: the UDF runs on Fused and the model receives the answer, not the dataset.
Fused's MCP surface: any Canvas in Workbench can be published as an MCP server, exposing the UDFs on it — Python functions that join and process data from files, cloud storage, databases and STAC catalogs — as callable tools.
- A ready-to-run connect command from the canvas itself: open it in Workbench, click Share, then Copy MCP
- Every visible UDF on the canvas discoverable and callable as a tool, with its parameters
- A visibility toggle per UDF node, so helper functions still run when called by other UDFs but stay out of the agent's tool list
- The same MCP URL working in Claude Code, the Claude desktop app, claude.ai, Cursor, VS Code and any other MCP client
- Compute that stays behind the endpoint — the agent gets the result, not the dataset
- An ai-chat widget you can drop onto the canvas itself, with access to every UDF wired into that widget node
A Fused account and a canvas in Workbench. Connection is over HTTP with a canvas token in the URL, added with claude mcp add --transport http. On Team and Enterprise plans only Owners can add the connector org-wide, after which each member connects and authenticates individually; the ai-chat widget requires viewers to be logged in to fused.io.
Build from source — clone the repository and build it, then point your client at the binary