Skip to content
Writ

Documentation

Connect an agent

Give a compatible agent access to selected Writ documents through one remote connection. Writ handles sign-in and consent in your browser; you never paste an access token into a conversation.

Connect

  1. 1. Add Writ as a remote connector

    In a client that supports remote MCP servers with OAuth, add https://writ.team/mcp/. This includes hosted agent products when they expose custom remote connectors; the setting may be called a connector, app, integration, or MCP server.

  2. 2. Sign in to Writ

    The client opens Writ in your browser. Sign in there. Your Writ password and session never pass through the agent.

  3. 3. Choose access

    Writ shows the client name, requested abilities, and your available documents. Select the documents it may use, then approve. Documents you do not select remain inaccessible, including private documents you can access yourself.

  4. 4. Ask it to work

    Ask the agent to read a document, start or continue an edit, autosave its work, submit it, discuss a finding, or review a submission. Writ attributes every action to the connected agent acting for your account.

The client stores the connection. Writ issues a short-lived access credential and a rotating refresh credential to the connector runtime, not to the language model. Reconnect through the browser if the client loses or revokes it. Never put a Writ credential in a prompt.

What the agent can do

Each connection receives only the abilities shown on the consent screen:

There is deliberately no agent tool that finishes a meaning-changing submission. Writ keeps that decision with a person. Comparison, blockers, approval policy, and document roles still apply to agent work.

Manage or revoke access

Open Connected agents from your account menu to see every connection, its documents, abilities, creation time, and recent use. Revoke access invalidates its access and refresh credentials and removes the connector principal from the selected documents immediately. Existing authorship and comments remain.

MCP and the optional skill

MCP is the live connection: it defines the tools, transports each call, and carries authorization. A skill is the workflow guide: it teaches an agent Writ's product language and when to read, autosave, submit, or ask for human action. A skill alone cannot securely reach your private documents.

The repository's Writ integration bundles both for clients that support plugins. Other compatible clients need only the remote MCP URL; their instructions can use the same workflow.

Local coding agents and the CLI

A local harness can use the same Writ skill without MCP. In that setup the skill calls the ordinary semgit CLI, while a separately provisioned developer token lives in the harness's environment or secret store:

SEMGIT_URL=https://writ.team
SEMGIT_TOKEN=<developer token>
export SEMGIT_URL SEMGIT_TOKEN
semgit edit start my-project --path doc.md
semgit edit get my-project edit-1 doc.md

The skill is still only instructions. The CLI sends the token as an Authorization: Bearer header, refuses to send it over remote cleartext HTTP, and redacts it from errors. Never put the token in a prompt, command argument, repository file, or transcript.

Create the token on Connected agents, selecting its documents and abilities. Writ shows the secret once; rotate it if it is lost and revoke it when the device or harness should stop working.

The REST CLI supports the complete read, edit, autosave, submit, review, finding, conflict, comment (semgit issues), precheck, and export workflow. For ordinary writing, use semgit edit. HTTPS Git is optional for bulk or offline repository work; configure a credential helper and use the developer token as the password. Hosted OAuth tokens are not Git passwords.

For raw HTTP endpoints and exact request shapes, see The API.