← Browse use cases · Hosting & Model Providers
| At a glance | Details |
|---|---|
| Outcome | Authentication patterns for supported Copilot, Codex, and Claude subscription backends. |
| Requires | Eligible coding-tool account; provider-specific OAuth or CLI setup; any required extra usage. |
| Available | Guide |
| Validation | Sources linked below; workflow not run during this catalog redesign. |
Class: Official integration · Confidence: High · Demo status: Config patterns
You already pay for GitHub Copilot, OpenAI Codex, or a Claude Max plan with extra-usage credits. Standing up Hermes on top of yet another API spend feels wasteful. You want Hermes to authenticate against the coding-tool subscription you already have — OAuth where the tool uses OAuth, API key where it uses a key — and to recover gracefully when a token expires mid-run.
Hermes ships first-class provider IDs for the three major coding-tool backends, each matching the subscription's native auth method instead of forcing a separate API key purchase.
- GitHub Copilot (direct API) — provider
copilot. Auth chain:COPILOT_GITHUB_TOKEN→GH_TOKEN→GITHUB_TOKEN→gh auth token. Accepts OAuth (gho_), fine-grained PAT (github_pat_), and GitHub App (ghu_) tokens — not classic PATs (ghp_*). On HTTP 401 it does a one-shot credential recovery (re-resolves token, rebuilds client, retries once). GPT-5+ models auto-use the Responses API; others use Chat Completions. - GitHub Copilot (ACP backend) — provider
copilot-acp. Spawns the local Copilot CLI as a subprocess and speaks ACP to it. Requires the Copilot CLI in PATH and an existingcopilot login. Overridable viaHERMES_COPILOT_ACP_COMMAND/HERMES_COPILOT_ACP_ARGS(defaultscopilot/--acp --stdio). - OpenAI Codex (device-code OAuth) — provider
openai-codex. Authenticates by device code (open a URL, enter a code). No Codex CLI install required. Credentials land in~/.hermes/auth.jsonand can be imported from~/.codex/auth.json. On refresh failure (HTTP 4xx /invalid_grant), marks the refresh token dead; re-login withhermes auth add codex-oauth. - Anthropic / Claude Max "extra usage" — provider
anthropic(aliasesclaude,claude-code). Three auth methods: API key (ANTHROPIC_API_KEY, pay-per-token); OAuth viahermes model(only works on a Claude Max plan with purchased extra-usage credits); setup-token (ANTHROPIC_TOKEN, legacy/fallback). Hermes prefers Claude Code's own credential store over copying the token into~/.hermes/.envand auto-detects existing Claude Code credential files.
# easiest: let Hermes reuse your gh login
gh auth login # once, if not already# ~/.hermes/config.yaml
model:
provider: "copilot"
default: "gpt-5.4"hermes model # → "OpenAI Codex" → device-code flow in browser
# or import an existing Codex login:
# credentials copied from ~/.codex/auth.json into ~/.hermes/auth.jsonIf the refresh token dies later:
hermes auth add codex-oauthhermes model # → "Anthropic" → OAuth (needs Claude Max + extra-usage credits)
# or API key:
# export ANTHROPIC_API_KEY=sk-ant-...# ~/.hermes/config.yaml
model:
provider: "anthropic"
default: "claude-sonnet-4-6"- An active Copilot subscription (direct) or Copilot CLI +
copilot login(ACP mode) - A Codex / OpenAI coding-plan account for the device-code flow
- For Claude Max: the Max plan with purchased extra-usage credits (plain Max without extra usage won't OAuth); or an
ANTHROPIC_API_KEYfor the pay-per-token path
- Token-type restrictions are real. Copilot's direct API rejects classic
ghp_*PATs — use a fine-grained PAT or an OAuth/App token. Don't waste time debugging a 401 on a classic PAT. - The 401 one-shot recovery on Copilot means a transient auth blip won't kill your session; a persistent 401 means the token is actually invalid and you re-
gh auth login. - Claude Max OAuth is conditional — only plans with extra-usage credits. If
hermes modelOAuth fails and you're on plain Max, fall back to an API key. - ACP-mode Copilot vs direct-API Copilot differ in process model: direct API is an HTTP client; ACP spawns the local CLI subprocess. Direct API is simpler; ACP is useful when you want the CLI's own session/auth handling.
- Mid-session switching (
/model copilot:gpt-5.4,/model anthropic:claude-sonnet-4-6) lets you spend cheap tokens for routine work and reserve your subscription's quota-billed models for the hard parts.
- AI Providers docs (rendered, Copilot / Codex / Anthropic sections): https://hermes-agent.nousresearch.com/docs/integrations/providers/
- Source-of-truth markdown: https://github.com/NousResearch/hermes-agent/blob/main/website/docs/integrations/providers.md