OpenRouter
OpenRouter gives Chat and Work access to hosted models when your hardware is not the right tool. The API key stays on this PC, encrypted; the rest of Studio — sessions, projects, tools and permissions — behaves the same.
What OpenRouter adds
- A large catalogue of hosted models without downloading weights.
- Streaming chat completions and OpenAI-style tool calls.
- A fallback when a local model is too large, too slow or not tool-capable.
- No local GPU or Compute host required.
It is inference-only. You cannot fine-tune an OpenRouter model from Studio, and your prompts leave this PC. For private data, use a local model or your own Compute host.
Connect it
- Create an API key at openrouter.ai/keys.
- Settings → Providers → OpenRouter. Paste the key and save. The base URL is already set to the OpenAI-compatible endpoint; only change it for a compatible gateway.
- Load models. Studio fetches the active model list, including free models. Press the refresh control if the list is stale.
- Pick a chat model. The provider picker on the welcome and chat screens lists OpenRouter models alongside local ones.
The key is encrypted with the same local key box as the Compute and HuggingFace secrets. It is never written into chats or exported model files.
Use it in Chat
- Choose OpenRouter in the provider picker, then a model.
- Streaming, stop generation, saved sessions and system prompts work as with a local model.
- Model badges and "won't fit" checks do not apply; the model runs remotely.
- Tool calls use the model's OpenAI-style tool support. Studio's tool-call recovery setting (Strict, Lenient, Loose) controls how forgiving it is when a model emits malformed calls.
- Free models may be rate-limited or unavailable at peak times; the error is shown in the chat.
Use it in Work
Work can use an OpenRouter model for reasoning while the project's files and tools stay local. The permission gate still applies:
- Context locations are explicitly added with read, write and delete permissions.
- Web access is a per-project permission.
- Skills and memory come from the project folder, not the provider.
- The send action is gated when the tool probe failed, so a misconfigured model cannot silently run tools.
Cost and privacy
- OpenRouter bills per token on your account. Studio does not add a margin.
- Free models are useful for evaluation; read their data-use terms.
- Prompt, completion and tool payloads go to the provider; the API key stays encrypted locally.
- For regulated or private data, prefer local inference or a Compute host you control.
Problems
| Message | Fix |
|---|---|
| 401 / invalid key | The key was revoked or pasted with whitespace. Create a new one and save it. |
| 402 / insufficient credits | Add credit or switch to a free model. |
| Model not found | Refresh the model list; the provider may have retired the id. |
| Rate limited | Wait or choose another model. Free tiers are shared. |
| Tool call not executed | The model may not support tools, or the tool probe failed. Check the project permissions and try Strict recovery. |
| Slow first token | The provider is queuing the request; a smaller model or another provider route is faster. |