Copilot CLI
Use Tokamak as Copilot CLI's Responses provider while retaining GitHub tools and authentication.
Install and launch
Complete Install and connect, substituting tokamak-stag on staging. Tokamak installs @github/copilot through npm if missing. GitHub authentication remains necessary for features that require it. Choose a model with Responses and compatible tool support.
tokamak launch copilot --model YOUR_MODEL_ID --check
tokamak launch copilot --model YOUR_MODEL_IDFor a headless prompt:
tokamak launch copilot --model YOUR_MODEL_ID -- -p "explain this repository"Use tokamak launch copilot --model for the catalog picker. Arguments after -- belong to Copilot. Claude-only launch options are rejected.
What changes
The child process receives the Tokamak base URL, key and selected model through Copilot's custom-provider environment. It selects the OpenAI provider type with the Responses wire API. Existing GitHub authentication, session files, permissions, MCP and plugins remain available.
This configures model inference. It does not replace GitHub authentication for repository or hosted features, or guarantee that every model supports Copilot's requested capabilities.
Usage and troubleshooting
Server metering records routed requests. With a resolved Tokamak user identity, the launcher configures Copilot's traces/metrics exporter; message content is captured only when your organization turned on Insights prompt summaries. If identity cannot be resolved, that supplemental exporter stays off; server metering still applies.
Copilot inference attribution may appear as unknown even when client activity has a Copilot label. Treat the Tools view's coverage information separately from request totals.
The recorded Windows check reached Tokamak Responses with Copilot 1.0.83 but received a 422 output-bound refusal. Copilot returned zero despite the failed request, so inspect the response and server usage rather than only its exit status. For output bounds, authentication or credit issues, follow troubleshooting.