Configure Copilot CLI

Install

The agent image includes copilot from the @github/copilot package, at the version pinned by agentVersions.copilot. Configure account access in your TypeScript harness, as shown below.

Your GitHub account needs Copilot access; see GitHub’s CLI quickstart.

Sign in with your account

Copilot always runs on your Copilot plan. Pass a fine-grained GitHub token with the Copilot Requests permission:

import { createAgent, createCopilotHarness } from "@elie-laloum/outpost";

export const coder = createAgent({
  harness: createCopilotHarness({
    authentication: { account: { variable: "COPILOT_GITHUB_TOKEN" } },
    variables: { COPILOT_GITHUB_TOKEN: process.env.COPILOT_GITHUB_TOKEN ?? "" },
  }),
});

To reuse a host login instead, run copilot login and set authentication: "account". Outpost reads the token of the last logged-in user from ~/.copilot/config.json (or $COPILOT_HOME/config.json) and passes it to the sandbox as COPILOT_GITHUB_TOKEN.

Caution

Copilot stores its login in the system keychain by default, which Outpost never reads. If config.json holds no token, use the token form above.

Other credential forms and where credentials go: Authentication.

Use an API key

Copilot has no API-key mode: authentication: "usage" is rejected when the agent is composed. Requests count against your Copilot plan.

Available features

• ConversationsCaptured as a session bundle, then resumed warm or cold with --resume.
• Typed responsesAn invalid answer is repaired by resuming the same conversation.
• SteeringOutpost stops the process and resumes the session with your text.
• MCP serversPassed on each run with --additional-mcp-config.
• Usage reportingToken counts read after the run from session-state/<id>/events.jsonl.
• Quota pausesCopilot rate-limit and credit messages fail with code quota.

API reference: CopilotSettings and createTransportConversations.

Limits

• No automated fork: Forking a Copilot conversation is rejected; resume it instead.
• For supported model settings, see CopilotSettings.
• Classic tokens: ghp_ personal access tokens are rejected.
• Incomplete usage: A missing or unreadable session file sets usage.complete to false, with a warning.
• Late counts: Token totals can arrive after the model has spent them. Pair a budget with a timeout.
• Premium requests: Copilot bills premium requests; usage reports tokens, not premium requests.
• MCP options: startupTimeoutMs and oauth are rejected; authenticate HTTP servers with bearerTokenVariable.

API: createCopilotHarness · CopilotSettings · createCopilotConversations.