You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Serve the agent's workflow tools from a Texera MCP server #9457
Today the agent's tools run inside the agent service and edit a private copy of the workflow. That copy is loaded once, before the first prompt, so a change the user makes between prompts gets overwritten. #9464 fixes this by moving the agent onto the shared workflow.
This task moves the tools from #9481 onto one Texera MCP server, with the same names and inputs. The agent and outside assistants (#8663) call it, and its tools edit the shared workflow through #9464. This revisits the inline-tools decision in #5610.
The server holds no session: every tool takes the workflow id, so any copy of the server can answer, as the current MCP revision (2026-07-28) expects.
It is built on the official server SDK (@modelcontextprotocol/server, Apache-2.0), which runs on Bun like the agent service.
Each tool says whether it only reads (readOnlyHint), so a client can ask the person before a change.
From aicam's prototype in Proposal: Texera MCP, let AI assistants work with Texera #8663, the shared client package and the test harness can be kept. Its workflow editing holds a session in the server's memory (workflow_open to workflow_save), which the shared session replaces.
The agent calls the server with the user's token, as its backend calls do today. The MCP specification, though, has a server accept only tokens issued for it, and today's Texera tokens name no audience. So an outside assistant can use a token the person pastes in, while signing in from an assistant needs an authorization server, which Texera does not have yet.
The dataset tools from #9377, which also cover #4588, will come later on the endpoints from #9395. Review mode (#9075) waits on #9464.
Activity