Claude Desktop

Statewright lives inside Code → Local in the ordinary Claude app. You do not need a second app, shortcut or Statewright-specific Claude executable. It does not add controls to ordinary Chat or Cowork.

Open your first conversation

1. Install the Claude plugin. Use Code engine 2.1.295 or newer; /status in a Code conversation reports the bundled engine.
2. Open Code, choose Local and select your project.
3. Send the first message. The blank composer has not started the Code session, so the Arbiter need not appear before this.
4. If you installed or updated the plugin with the conversation open, run /reload-plugins.

Upgrading a manual-hook installation? Complete the legacy-hook migration before opening a workflow. Updating the plugin does not remove old hooks from Claude settings.

Anthropic's mod documentation describes the host's supported surfaces and plugin controls.

Connect once, then check the result

Click the connection-status pill beside arbiter and statewright.ai ∞. This opens the sidebar; opening it alone does not sign you in. Click Connect and complete the native permission and Statewright browser sign-in flow. Keep the conversation open until its workflow list appears.

An existing login may be restored without opening a browser. Conversely, a browser success page alone does not prove the conversation has connected. The workflow list is the useful confirmation.

Claude owns the saved OAuth credentials. There is no Statewright API key to paste into the conversation. If the browser is unavailable, use the manual sign-in path.

Choose and start

In Choose a workflow, select a name, then click Start workflow. You can also ask Claude to use that exact workflow for your task. Both paths use the same conversation-bound integration.

Selection is not execution. A successful start shows the workflow's current state. If another workflow is already active, continue it or turn it Off before starting a different one. Repeating a request for the active workflow does not start a second run.

Read the Arbiter and results

A successful transition reads previous state => new state. A pending review stays pending; it is not rendered as a completed move. Reasoning is the receipt's stated rationale, not a transcript of a model's private reasoning.

The sidebar distinguishes the requested route from the last observed response. Effort is a requested setting. A model label alone is not proof of execution.

Review, finish or turn Off

When a workflow needs human review, choose Review this step and inspect the presented request. Cancel, dismissal or disconnect does not approve it. Claude's own permission to run a tool is separate from this workflow decision. Statewright does not manufacture an Always allow option if Claude offers only a one-time decision.

A final state completes the workflow; it does not mean every test passed unless the run's evidence says so. Turn Statewright Off deactivates enforcement and restores the host-selected model. Back to conversation only closes the pane.

Recover without losing context

• No pills or command: send the first message, check /status and /plugin, then reload plugins. Organization policy may prevent mods from loading.
• Connection needs attention: click it and read the failed step. Choose Try connecting again for a connection failure.
• Check Claude permissions: review the native permission prompt or rule. Do not keep signing in or grant blanket permissions.
• Missing tools after an update: quit and reopen Claude Desktop, then resume the conversation. Update Statewright if its tools still are not loaded.
• A state looks stale: use Refresh workflows before continuing. Do not restart a workflow merely because the sidebar was closed.
• A tool or command is denied: retain its message and current state. Follow the allowed tools/transitions, or ask the user whether to change the workflow.

See Claude enforcement limits.