Hermes Agent BYOK: Bring Your Own API Key

Why BYOK?

• No markup. Pay raw provider pricing — usually 10–30% cheaper than bundled inference at scale.
• Direct billing. Receipts come from your provider, not from Hermes hosting.
• Higher rate limits. If you already have a Tier 4 OpenAI account or a custom Anthropic agreement, you keep those limits.
• Existing relationships. Use enterprise agreements, committed-use discounts, or free-tier credits you already have.
• Data residency. Your provider, your data-handling agreement.
• One key, many models. A single OpenRouter or LiteLLM key unlocks 200+ models across all providers.

Supported BYOK Providers

Hermes also accepts OAuth credentials for some providers — including Claude (CLAUDE_CODE_OAUTH_TOKEN) and Codex (device-flow). These let you use a paid Claude Code or ChatGPT Plus subscription as the inference source instead of an API key.

BYOK on OpenClaw Launch (Managed)

1. Go to openclawlaunch.com/hermes-hosting.
2. In the configurator, toggle BYOK.
3. Pick a provider (OpenAI, Anthropic, Google, OpenRouter, Mistral, DeepSeek, xAI, Groq, Together, Cerebras, Z.ai, or DashScope).
4. Paste your API key. It's encrypted at rest with your account's master key.
5. Pick a model and deploy. 100% of inference flows through your provider account.

BYOK on Self-Hosted Hermes

On your own server, BYOK is just environment variables. Pick the provider you want and export its key:

# Pick one or more — Hermes routes by inference.provider
export OPENAI_API_KEY=sk-proj-...
export ANTHROPIC_API_KEY=sk-ant-...
export GOOGLE_API_KEY=AIza...
export OPENROUTER_API_KEY=sk-or-...

# Tell Hermes which one to use as default
hermes inference set anthropic
hermes model set claude-sonnet-4.6

# Or via /opt/data/config.yaml:
# inference:
#   provider: anthropic
# model:
#   default: claude-sonnet-4.6

You can keep multiple keys exported and switch providers at runtime with hermes inference set <provider> — useful when one provider has an outage or rate-limit issue.

One Key, Many Models: OpenRouter or LiteLLM

If you don't want to manage a dozen keys, OpenRouter gives you one key for 200+ models. Switch between Claude, GPT, Gemini, DeepSeek, Llama, and dozens more via Hermes's /model command without touching env vars.

export OPENROUTER_API_KEY=sk-or-...
hermes inference set openrouter

/model anthropic/claude-sonnet-4.6
/model openai/gpt-5.5
/model google/gemini-2.5-pro
/model meta-llama/llama-4-maverick

For enterprise BYOK with custom routing rules, point Hermes at a self-hosted LiteLLM proxy. Hermes treats it as an OpenAI-compatible endpoint, and LiteLLM handles the routing, caching, and fallback logic.

Cost Math: When Does BYOK Save Money?

For low-volume use, the bundled AI credits on OpenClaw Launch are usually cheaper than BYOK once you account for provider minimum spend, pre-paid credit fees, and the operational overhead of managing keys. BYOK starts winning around 500–1,000 messages per day, where the unmarked-up inference cost more than offsets the credits the plan includes.

Heavy users (10K+ messages/day) typically save 20–40% on inference with BYOK vs bundled. For coding agents on Claude Sonnet, the savings can be much larger if you use a Claude Code subscription via CLAUDE_CODE_OAUTH_TOKEN.

Security

On OpenClaw Launch, BYOK keys are encrypted with your account's wrapped master key — the same envelope that protects all your saved configs. Keys are decrypted only inside your container, never in the orchestrator. On self-hosted Hermes, keys live in your /opt/data/config.yaml or environment; secure that file as you would any other credential.

What's Next?

• Hermes + Claude — BYOK setup for Anthropic
• Hermes + OpenAI — BYOK setup for OpenAI
• Hermes + OpenRouter — One key for 200+ models
• Hermes + MCP — Add tools your BYOK agent can use