MCP
RequestZ ships an embedded MCP server. Claude, Cursor and any MCP client can create and edit requests, run collections and flows, read linter problems and push changes to git — talking to YOUR app on YOUR machine, never to a cloud.
Download RequestZTwo paths: direct stdio (Claude Code, Codex) or HTTP with a credential (Cursor and friends).
claude mcp add requestz -- requestz mcp
No tokens to copy: the bridge reads the rotating credential from disk — the same trust model your terminal already uses.
http://127.0.0.1:<puerto>/mcp Authorization: Bearer <credencial>
In the app: ⚙ → MCP → “Generate credential”. The token rotates on every server start.
What the AI can do is exactly what you can do in the app — no more, and never through a side door.
Create, edit, move and document — with apply_batch: one transactional batch of changes with a SINGLE approval, all or nothing.
Run a request, a collection or a flow and read results, asserts and tests — real sends always ask permission.
Read the Problems panel (plaintext credentials, duplicates, OpenAPI drift) and fix what it reports.
Pull, status and sync (commit + push, always authorized) — the AI can push ONLY the files it touched.
Read and write variables per scope; vault names yes, vault secrets never.
Import cURL and OpenAPI, diff collections against the spec, and drive the mock server.
What no other API client can close today:
Your assistant works on the team repo with the team's flow — and whatever it pushes lands as a diff anyone can review in a PR.
Designed so letting an AI in doesn't require faith.
The server lives inside the app. AI↔workspace traffic never touches a server of ours — there isn't one.
Sending requests, writing variables, committing: every sensitive action is approved in the app, by hand.
The HTTP token rotates on every start; the stdio bridge reads from disk, like your terminal.
Anything that looks like a credential leaves the MCP masked. The vault is never written over MCP.
Any MCP client: Claude Code and Claude Desktop over stdio, Cursor and others over HTTP. MCP is an open standard; if your tool speaks MCP, it speaks to RequestZ.
No. Vault values never leave over MCP (only variable NAMES), and anything credential-shaped travels masked. Sends that use secrets are executed by the app, not the AI.
Yes: MCP is part of Pro, like everything that automates. Reading and sending requests by hand is free forever.
Every change is a file in git: it shows up in a diff, gets reviewed in a PR and reverts with git. That's why RequestZ's MCP can be bold — the repo is the safety net.