SORVEL/ DOCSGET STARTED
  1. Sorvel
  2. Docs
  3. Coding agents (MCP)

CODING AGENTS (MCP)

Sorvel is also an MCP server: Claude Code, Cursor, VS Code and Codex can drive it from their own chat. It runs in two places — inside the desktop app, next to the projects on your own machine, and on sorvel.dev for projects that live there. Either way Settings → MCP shows the address and a ready-made setup for each client (one command for Claude Code, one-click buttons for Cursor and VS Code, a TOML block for Codex). On sorvel.dev you paste nothing: the client opens this site, you approve once — reading, writing and spending credits are three separate permissions — and you withdraw it under “Connected applications”. The desktop app has one machine token instead, and access tokens are still there on sorvel.dev for CI and machines that cannot open a browser.

WORKFLOWAsk the agent to call sorvel_whoami to check the connection, then sorvel_get_project: it returns the file contract (edit .ts/.tsx under app/ and src/; the listed locked files are Sorvel’s) and, in the desktop app, the project folder on your disk. The agent edits with its own tools there, or with sorvel_write_files on sorvel.dev, then calls sorvel_validate — the same prettier, eslint, tsc and expo export Sorvel runs on generated code — which republishes the preview, refreshes your phone and, with a message, takes a checkpoint.
DELEGATINGsorvel_send_message is the chat turn itself: same model, same credits, charged only on success. It returns a job id at once and the agent polls sorvel_get_job; a clarifying question from Sorvel arrives in the result and is answered with another message in the same conversation. dry_run shows the price first. Everything else — files, history, checkpoints and restore, the phone preview — is free.
WHAT IS NOT THERENo tool runs shell commands, reveals secret values, touches integration tokens or publishes to a store; those stay behind the buttons in the app. An agent acts as you and never more: it sees only your projects, only the permissions you approved, and you end that from the same page — instantly. A write is refused outright if a file breaks the contract or carries a secret. Codex gives a tool 60 seconds by default, so long work is always a job you poll.

Seventeen tools, grouped by what they are for. Only sorvel_send_message costs credits; everything else is free:

  • Finding your way around — sorvel_whoami (who is signed in, and whether the engine is ready), sorvel_list_projects, sorvel_get_project — the file contract, what is locked, the preview address, the last checkpoint — and sorvel_create_project.
  • Reading and writing the code — sorvel_list_files, sorvel_read_file, sorvel_write_files and sorvel_delete_file. In the desktop app the agent can skip these and edit the folder with its own tools; on sorvel.dev they are the only way in, and a write is refused outright if it breaks the contract or carries a secret.
  • Checking the work — sorvel_validate runs prettier, eslint, tsc and expo export, republishes the preview and refreshes the phone, and commits a checkpoint when you give it a message. sorvel_device_preview starts, stops or inspects the phone preview and can return the QR image.
  • Handing work to Sorvel — sorvel_send_message is the chat turn itself, the one tool that spends credits (dry_run prices it first); sorvel_get_job follows it, sorvel_cancel_job is the Stop button, and sorvel_get_messages reads what Sorvel said or asked.
  • Going back — sorvel_list_checkpoints and sorvel_restore_checkpoint, which restores as a forward commit, so the history it came from is still there.

Settings → MCP carries the exact setup for your client, and the same list with one line per tool.

Coding agents and the MCP server — Sorvel Docs