| name | add-custom-models |
| description | Enable custom model support — use local models (OMLX, Ollama) or any third-party model endpoint (HuggingFace, Together AI, Groq, etc.) alongside your primary agent SDK. Triggers on "custom models", "local models", "add custom", "add local". |
Add Custom Models
Enable support for custom model endpoints — local models, third-party providers, or self-hosted deployments. Uses the Codex SDK as the bridge to any OpenAI-compatible endpoint.
Phase 1: Pre-flight
Check if Codex SDK is installed
ls src/runtime/codex-runtime.ts 2>/dev/null && echo "CODEX_INSTALLED" || echo "CODEX_MISSING"
If CODEX_MISSING, the Codex SDK is required as the bridge for custom models. Invoke /add-agentSDK-codex first, then return here.
Check current state
grep -q 'OMLX_URL\|LITELLM_URL' .env 2>/dev/null && echo "ENDPOINTS_CONFIGURED" || echo "NO_ENDPOINTS"
Phase 2: Configure
Explain to the user
Custom models let you use any OpenAI-compatible endpoint alongside your primary agent SDK.
This works by routing through the Codex SDK — when you select a custom model via /model, the system automatically uses Codex with the custom endpoint URL.
Common endpoints:
• OMLX — local models optimized for Apple Silicon (persistent KV cache, tool calling)
• Ollama — easy local model hosting
• Together AI, Groq, Fireworks — fast third-party inference
• HuggingFace Inference — hosted open-source models
• Self-hosted vLLM, llama.cpp — your own deployment
Run /add-model-endpoint to add endpoints one at a time.
Update available models in config
After running /add-model-endpoint for each desired provider, verify models appear in /model command output.
Rebuild container
Only needed if this is the first time installing custom model support:
npm run build && ./container/build.sh
Restart service
launchctl kickstart -k gui/$(id -u)/com.nanoclaw
Phase 3: Verify
Send /model in Telegram. You should see your custom models listed below the cloud models.
Switch to a custom model:
/model omlx/llama-3.1-8b
Send a message and verify it responds using the local/custom model.
Switch back:
/model gpt-5.4-mini
Removal
Remove endpoint URLs from .env and remove the custom model entries from AVAILABLE_MODELS in src/config.ts. Rebuild and restart. The Codex SDK can remain installed.
