Proposal: Texera MCP, let AI assistants work with Texera

[aicam]

Feature Summary

Many people now use AI assistants such as Claude, ChatGPT or Cursor for their daily work. Today those assistants cannot see or use Texera, so users have to switch back and forth and copy things between the two by hand.

This proposes a Texera MCP: a small connector that lets any AI assistant work with a user's own Texera account. A user could simply ask:

• "Upload this CSV as a new dataset."
• "Build a workflow that cleans this data and trains a classifier."
• "Run my workflow and tell me what the results look like."

The assistant would do the work in Texera, and the user sees the result in the normal Texera workspace.

Proposed Solution or Design

• The user connects their assistant to Texera by giving it the address of their Texera deployment and their login token. It works with any deployment, not only one hosted site.
• The assistant acts as that user, with the same permissions they already have. It cannot see anything the user cannot see.
• Ideally, changes appear live in the workspace, so the user can watch the assistant edit a workflow, step in, or undo.
• Everything the assistant builds is an ordinary Texera workflow or dataset that can be opened, shared and edited as usual.

A working prototype is in aicam/texera, branch feat/texera-mcp. It is meant to show the idea, not as a final design.

Questions for the community

1. Would you find this useful, and for which tasks?
2. Should it be part of the main Texera repository?

Affected Area

Other (AI assistant integration)

[sshiv012]

I like the idea of letting assistants work directly with Texera. Could a first step be exposing and documenting the relevant APIs/tools, or a CLI, with MCP as a wrapper later? That would also help with scripting and maintenance.

My hesitation is the additional code to maintain, potentially another running server, and the context consumed by tool definitions. I’m also unsure how MCP will hold up long term given how quickly AI conventions change.

From my experience working on the AsterixDB MCP over the summer, newer models were already quite good at calling REST APIs/tools with little to no documentation. Starting with a reusable API/tool layer seems useful either way, while keeping the door open for MCP.