Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Teams

Retool's MCP server is generally available

Written November 13, 2026

Retool made its MCP server generally available. What it lets an outside assistant do with your Retool apps, per Retool's own material.

Out of beta

Retool's MCP server announcement now opens with the status change: the server is out of public beta and generally available. Retool says thousands of builders used it during the beta while capabilities were added. For a team that waits for a GA label before connecting an outside assistant to its internal-tools platform, that label is the news.

The additions Retool names alongside GA are administrative. Admins can monitor MCP token spend in the AI monitoring dashboard. The toolset has grown to cover updating access to groups, roles and permissions, and the server can inspect and update permissions on workflows as well. Governance tooling arriving with the GA label, rather than after it, is the shape a platform team wants to see.

What an assistant can do through it

Retool's stated purpose is to let builders work on Retool from the coding environment they already use — Claude, Cursor, Codex, ChatGPT and Kiro are the clients its setup material names — instead of switching to a browser. The tool calls it lists include:

That is a wide door: app content, data queries, and user administration behind one connection. Retool's own framing is that the server extends the platform's security model to the connected agent — changes made through it inherit the access controls, audit trails and data permissions the organization already set at the platform level, rather than being configured per app or per prompt.

How the connection works

The connection details are short. The server is reached over HTTP at an address on your own Retool instance, in the form https://<your-retool-instance>/mcp, and authentication is OAuth 2.0. After authenticating, Retool notes, you may need to restart your coding agent. The server is available for cloud and self-hosted customers.

Because access runs through OAuth sign-in as a user, what an assistant can reach follows that user's standing in the organization. The GA additions above give admins their side of the picture: spend monitoring in the AI dashboard, and permission management — for groups, roles and workflows — available through the same server.

The production-data question

Retool followed the GA note with an engineer's account of using the server as a gateway to production data, which is where the governance framing stops being abstract. An assistant that can query connected resources and edit the apps built on them is working close to live systems. Retool's position is that the platform's existing controls — who can see a resource, what an app may do with it, what gets audited — are the right place for that risk to be managed, because they apply no matter which client made the change. A team evaluating the GA server should test that claim against its own setup: connect as a restricted user first, and confirm the assistant's reach matches that user's reach and nothing beyond it.

What to do

Related playbooks

Sources

Back to: Teams playbooks hub · Also: Run Claude Managed Agents in your own cloud