Claude Cowork

Before you start

Install Claude Desktop on macOS or Windows and complete Install and connect. Use tokamak-stag for staging. Your deployment needs an eligible Claude-capable model with Messages, streaming and tool support.

Configure and open

tokamak launch claude-cowork --check
tokamak launch claude-cowork

The first command checks local configuration offline. The second saves a Tokamak third-party deployment profile and opens Claude. Fully quit and reopen Claude if it is already running; active tasks are not terminated.

Choose your model inside Claude. Cowork fetches eligible models from the configured Tokamak gateway's authenticated /v1/models endpoint. The launcher rejects --model and does not apply your saved CLI model default.

Model discovery

For Tokamak discovery, use recognizable Claude model IDs backed by a compatible Claude route. Tokamak's catalog does not publish the optional family-tier override supported by Claude Desktop. A non-Claude model in the general catalog does not automatically become a Cowork model. Tokamak does not invent a fallback or force a context size; it lists each model's max_input_tokens, which is how Claude offers a model's 1M context.

If the picker is empty, check authentication and ask your deployment administrator to confirm an eligible Claude-capable route. If an old conversation retains an unavailable model, select an available model in Claude or start a new task.

What changes

Tokamak manages a named profile in Claude-3p/configLibrary and selects it; current Claude Desktop (2.x) opens its third-party deployment because that profile is applied. Tokamak also records the deployment mode in the Claude and Claude-3p desktop configurations. This affects the regular Claude desktop app too. If you switched Claude back to Claude.ai sign-in, the next launch says so and switches it to Tokamak again. Other profiles, MCP configuration, permissions and organization-managed policy remain in place.

The profile contains your API key in a private file. Keep its backups and restore journal private. The launcher enables model discovery and removes its older fixed model list. It does not enable permission bypass, auto mode or unrestricted network access. Client activity export depends on server support; prompt text is captured only when your organization turned on Insights prompt summaries, and inference metering remains separate.

Restore your previous setup

tokamak launch claude-cowork --restore

Restore works offline and checks for changes before restoring files. Fully reopen Claude afterward. Formatting-only JSON changes are accepted, and unrelated desktop preferences and MCP settings are preserved. Profile, selection metadata and deployment-mode changes require reconciliation with the original .tokamak-backup files. Rerun the launcher after authenticating with a new account or rotating a key.

Troubleshooting and limits

• If an older profile returns to standard sign-in, rerun the updated launcher to migrate its previous profile to the current UUID-based format, then fully reopen Claude.
• For a nonstandard app executable, use an absolute TOKAMAK_CLAUDE_DESKTOP path. A custom TOKAMAK_CLAUDE_DESKTOP_DATA selects configuration files; it does not move the app's data directory.
• Subscription, hosted connectors and remote features retain Claude's own deployment requirements.

The launcher was last reviewed against Claude Desktop 2.19675.0 and supports 2.7032.0 or newer; a launch warns on an older build and never blocks. The recorded Windows check (Claude Desktop 1.46388.4) verified profile loading and catalog access, but the test catalog had no eligible Claude model. Successful Cowork inference is not established by that check. See troubleshooting.