Build Chamfer apps from Claude, Cursor or Codex

What your editor can do

The MCP server offers the same actions as the Bench, limited to your role:

• list the apps, data sources and environments you can open

• create an app from a prompt, or change one screen, query or permission

• run a query against staging and read the result

• open a release for review, with the diff attached

If your role can’t do something in the browser, it can’t do it over MCP either. A support agent’s editor can’t read the payments schema, however the request is worded.

In the private beta, most people used it for the last stretch of an app: a tricky query, a custom component, a rename across twelve screens. The first draft still tends to start from a prompt in the browser. The fine work moves to the editor, next to the rest of the team’s code.

Connect in two minutes

Every workspace has its own server URL. In Claude Code it’s one command:

claude mcp add chamfer --transport

claude mcp add chamfer --transport

claude mcp add chamfer --transport

Restart the client, type /mcp, choose chamfer and sign in with your company SSO. Cursor and Codex take a short config block instead. The docs have the exact steps for each.

Admins decide which clients are allowed. Under Settings → Connections you can allow all three, pick one, or switch MCP off for the whole workspace.

Plans and limits

MCP is on every plan. Edits made over MCP draw on the same monthly credits as the Bench, and running a query uses none.

Your name on every change

MCP sessions sign in as you, through your identity provider. There are no shared tokens to rotate and no service account with broad access. When someone leaves and their account is removed upstream, their session stops at the next call.

Each action lands in the Logbook with your name, the client you used and the prompt behind it. An auditor sees “Maya Okoro, via Cursor” next to the diff, not an anonymous API key. If a laptop goes missing, revoking its client ends every session it holds, and the record shows exactly what it did before you stepped in.

The connections panel shows every client signed in to your workspace. Admins can revoke any of them in one click.

Releases still go through a person

Working from an editor doesn’t skip review. A release opened over MCP joins the same queue as one opened in the browser, and the approver sees the same diff. If your Specs call for a second approver on payment writes, that rule holds whichever client wrote the change.

Approvers can also see where each edit came from. The release lists the client and the prompt behind every change, so a reviewer can tell a hand-written query from a generated one at a glance.

“Engineering stopped owning admin panels. We review the Specs, approve the permissions and get back to the product.”

Theo Marchetti, VP Engineering, Quenby

What we heard in the beta

About 600 workspaces used MCP in a private beta over the summer. Two requests shaped the launch. Engineers wanted to review from the editor too, so release diffs can now be read over MCP. Security teams wanted to know which client made each change, so the Logbook records it on every line.

What’s next

In October we’ll add Jobs to the MCP server, so you can schedule and test a run from your editor. We’re also building a read-only connection for teams that want AI clients to inspect apps without changing them. Tell us what else you need at hello@chamfer.dev.