Models & LLM Keys

ElasticClaw supports API-key providers and subscription-backed coding providers in the server UI and bootstrap path. Use these provider IDs in llm_keys[].provider and as the prefix in default_model.

Instead of a single API key, ElasticClaw uses named LLM keys inhub.yaml. Each key has a provider plus either an API key or an auth profile, and an optional default model. One key can be marked default: true. Configure more than one key when different workspaces should use different providers or accounts.

llm_keys:
  - name: anthropic-prod
    provider: anthropic
    api_key: ${ANTHROPIC_API_KEY}
    default: true
    default_model: anthropic/claude-sonnet-4-6

  - name: fireworks-kimi
    provider: fireworks
    api_key: ${FIREWORKS_API_KEY}
    default_model: fireworks/accounts/fireworks/models/kimi-k2p6

  - name: codex-chatgpt
    provider: codex
    auth_profile: codex-default
    default_model: codex/gpt-5.5

  - name: grok-build
    provider: grok
    auth_profile: grok-default
    default_model: grok/grok-build-0.1

The default_model field uses the provider/model format (e.g. anthropic/claude-sonnet-4-6).

Fireworks model choices in the hub are loaded from the Fireworks List Models API when a Fireworks key is configured. The examples on this page are representative only; the dropdown should show the currently available Fireworks models returned by your account.

If a model is not listed yet, choose the custom model option and enter the full Fireworks model ID, such as fireworks/accounts/fireworks/models/....

Codex and Grok Build are not configured with API tokens. They use the same subscription-backed account login that their coding CLIs use. Your provider subscription controls access, model availability, and usage limits; ElasticClaw only stores the resulting device-login auth profile.

1. Open Settings → Models in the ElasticClaw hub.
2. Add a model key and select Codex or Grok Build.
3. Enter a stable CLI auth profile name, such as codex-default or grok-default.
4. Click Login, open the device-login URL, and enter the one-time code shown in the hub.
5. After the profile is saved, choose the default model and save the model key.

The hub starts a provider device-login flow and shows the URL plus a one-time code. After you approve the login in the provider's browser flow, the hub stores an opaque model_auth_profiles entry and links it from the named llm_keys entry with auth_profile.

llm_keys:
  - name: codex-chatgpt
    provider: codex
    auth_profile: codex-default
    default_model: codex/gpt-5.5

model_auth_profiles:
  - name: codex-default
    provider: codex
    mode: device
    auth_state: <managed by the hub login flow>

You normally should not hand-edit auth_state. During claw bootstrap, ElasticClaw restores the saved profile into the sandbox in the format expected by the selected provider CLI.

Set a global default model at the server level. All agents use this unless overridden.

default_model: anthropic/claude-sonnet-4-6

Workspaces can override the default model and specify which LLM key to use:

# elasticclaw-config.yaml
default_model: fireworks/accounts/fireworks/models/kimi-k2p6
llm_key: fireworks-kimi

If llm_key is set and default_model is empty, ElasticClaw Server resolves the model from the named key's default_model. If nollm_key is set, ElasticClaw Server uses the key marked default: true, then the server-level default_model.