GitHub Copilot

The copilot provider uses native GitHub authentication and provides direct access to Anthropic and OpenAI models via Messages or Responses wire formats respectively. The Responses API supports websockets.

The native provider sends Copilot-Integration-Id: copilot-sdk by default on Messages, Responses HTTP/SSE and WebSocket requests.

Setup

uv pip install fast-agent-mcp
fast-agent auth provider login copilot
fast-agent --model copilot.claude-sonnet-5

Alternatively, you can use fast-agent go and login direct from the model selection screen. When signed out, pressing Enter on a Copilot model starts the device-code login immediately. Follow the displayed GitHub URL and code; fast-agent waits for approval. Press Ctrl+C to cancel.

NB: OAuth credentials are not model entitlement: your account and application still need access to the selected model. Routing uses the local curated model registry, not the remote /models catalog. The fast-agent public OAuth client ID is Ov23li9BBH5sVoKopuI6; no client secret is required.

Credentials use shared provider-auth storage: the OS keyring when writable, otherwise ~/.fast-agent/auth.json. Set FAST_AGENT_AUTH_FILE to use an explicit portable provider credential file. Manage the stored Copilot credential with:

fast-agent auth provider show copilot
fast-agent auth provider token copilot
fast-agent auth provider export copilot ./copilot.auth.json
fast-agent auth provider logout copilot

show reports credential status. token prints a secret and export writes a secret-bearing file: keep both out of chat, logs and version control.

Headless authentication

Supply either an exported provider file via FAST_AGENT_AUTH_FILE or an appropriately authorized COPILOT_GITHUB_TOKEN from your secret manager. COPILOT_GITHUB_TOKEN takes precedence over stored credentials; GH_TOKEN and GITHUB_TOKEN are ignored. A rejected environment credential never starts interactive login or silently falls back to stored credentials. Noninteractive inference does not launch login.

Configuration

Optional defaults in fast-agent.yaml, generated from CopilotSettings:

copilot:
  base_url: 'https://api.githubcopilot.com'
  integration_id: 'copilot-sdk'
  runtime_timeout_seconds: 30.0
  cache_mode: 'auto'
  cache_ttl: null

• base_url: HTTPS origin, default https://api.githubcopilot.com. No path, query, credentials or fragment allowed.
• integration_id: configurable routing identity, default copilot-sdk. Validated as a safe HTTP header value and applied consistently across native transports. Configure in YAML or with COPILOT__INTEGRATION_ID for a custom ID; incidental request-header overrides cannot replace it. A valid value does not imply GitHub registration or model access.
• runtime_timeout_seconds: positive timeout for runtime operations.
• cache_mode: Claude Messages caching policy: auto caches the stable prompt and advances through recent conversation turns; prompt caches tools, system instructions and prompt templates; off disables automatic markers.
• cache_ttl: optional Claude Messages TTL override. null preserves the model default; explicit values are 5m and 1h. One-hour caching has not been live verified through Copilot. These cache settings do not affect GPT Responses.

When migrating older configuration, rename direct_base_url to base_url and remove backend, use_environment_token and cli_path. These obsolete settings are rejected, not silently ignored. Remove COPILOT__BACKEND from runner configuration too.

Exact model names

Use these provider-qualified names; there are no Copilot aliases. Unknown IDs are rejected rather than routed to another provider. copilot.claude-opus5 is not canonical and has no alias; use copilot.claude-opus-5.

Provider identity stays Copilot regardless of wire API; Anthropic/OpenAI API keys are not used for these models. The status bar prefixes Copilot model labels with (cp); this does not change the model name used in configuration.

Model parameters

Copilot-routed models inherit the base model's local metadata parameters, so copilot.claude-opus-5 gets the same model_specific prompt text, edit-tool contract, poll settings, structured-output policy and tokenizer support as anthropic.claude-opus-5. Only route-specific fields differ:

• default_provider, response_transports, response_websocket_providers: set from the curated Copilot model spec, not a remote catalog. Gateway acceptance is checked only by actual inference.
• response_service_tiers: cleared; Copilot rejects service-tier selection.
• codex_responses_lite: off; this is a Codex-only request contract.
• long_context_window: cleared; extended-context tiers are an upstream billing feature the gateway does not expose.
• anthropic_web_search_version, anthropic_web_fetch_version, anthropic_required_betas: cleared; the gateway exposes no Anthropic server-side web tools or beta headers.

Prompt caching

Claude Messages uses the normal Anthropic automatic cache planner by default: markers cover the tools/system prefix and advance through recent conversation turns while retaining the previous cache boundary. No manual cache_control metadata is needed. The TTL is inherited from the resolved model unless copilot.cache_ttl overrides it; missing TTL metadata falls back to the standard Anthropic settings default, not your Anthropic account configuration.

Use copilot.cache_mode to change this policy. Direct anthropic.cache_mode, anthropic.cache_ttl and diagnostic settings do not affect Copilot. Explicit request cache controls are still forwarded even when automatic planning is off.

Transport and compatibility

The four GPT models default to WebSockets, using static curated capability metadata; no remote catalog is required. Select HTTP/SSE explicitly if needed:

fast-agent --model 'copilot.gpt-6-astra?transport=sse'

The default (and explicit ?transport=websocket) has no SSE fallback. Explicit ?transport=auto tries WebSocket first, with SSE fallback on early errors. Ordinary OpenAI/Codex Responses defaults to auto. Claude models use SSE only. The native runtime does not query /models.