Coda has an assistant built in. The MCP server is the same thing from the other side: point Claude, ChatGPT or any other MCP client at one URL, ask for a pipeline in plain language, and get back a link that opens the finished workflow in your browser.
MCP — the Model Context Protocol — is how an AI client learns about a tool.
It reads Coda’s own node catalogue, adds the nodes, connects the ports and sets the parameters — then checks the result against the same rules the cards on the canvas use, and fixes what it finds before answering.
The workflow opens in Coda in your browser, where your credentials are, and you run it, edit it, and take it from there. Nothing is computed on the way.
The server reads the node definitions out of the deployed build, so a node added this week is one the model can use today — no version to match, nothing to upgrade.
^LC4$ →
Connectivity, downstream → Group By postType → Bar Chart. Open it and press Run:
It is hosted, so there is nothing to install and no key to paste. Wherever your client keeps remote MCP servers, add this URL:
One line in a terminal:
claude mcp add --transport http coda https://flyem.mrc-lmb.cam.ac.uk/coda-mcp/mcp
Anything that speaks MCP takes a remote server by URL — in a connector setting, or in the client’s own config file, as below.
Then just ask. A model that has never seen Coda reads coda_guide first, which is the whole catalogue — every node type, its ports and its parameters — so no prompting of your own is needed.
Add the URL to your client, then ask it for a connectome workflow. If you would rather not leave Coda at all, the assistant inside the app does the same job with your own API key.