v2.06
Sep 6, 2026
Latest
GitHub Copilot & Puter AI Providers, and a Resizable AI Panel
✦ New features
AI
- GitHub Copilot as an AI Provider — The assistant can now run on your own GitHub Copilot entitlement: an employer-granted seat, a personal plan or Copilot Free. There is no API key to create — pick Copilot in Preferences ▸ AI, sign in with the code the app shows you at github.com/login/device, and the model list fills in with exactly what your plan and your organisation's model policy allow. A GitHub Enterprise host can be entered in the Base URL field.
- Puter as an AI Provider — One token in front of roughly 900 models from Anthropic, OpenAI, Google, xAI, DeepSeek, Alibaba, Z.AI, Moonshot, MiniMax and Mistral, instead of registering with every vendor separately. A Puter account carries a free monthly allowance of credits, spent at each upstream vendor's own rate; Puter does not publish its size, and anything past it is billed to your own Puter account. The model list, the context gauge and the running cost figure all come from Puter's public catalogue, so the models and their rates are visible before you have a token at all.
- Grok as an AI Provider — Added xAI's Grok models, using your own xAI key. Paid — xAI publishes no free tier.
🛠 Fixes & improvements
AI
- The AI Panel Can Be Made Narrow — Dragging the AI side panel in stopped roughly halfway, and the point it stopped at moved while a turn was running, which is what made it look intermittent. The provider and model chips, the per-turn toggles, the session strip and the status line now wrap or shorten their labels instead of each demanding its full width, so the panel narrows to about a third of its old floor and hands the space back to the terminal.
- GitHub Models Removed — GitHub shut its GitHub Models service down on 30 July 2026 and the endpoint now refuses every request, so no key can bring it back. It is gone from the provider list. An installation that still had it selected now falls back to a provider that is configured and says on screen that it did so and why; your saved choice and key are left untouched.
- A Disabled Connector Kept Sending Its Tools — Switching an MCP connector off, or removing it, stopped the server but left its tools advertised to the model until the app was restarted. Two large connectors with one of them disabled was enough to push a request past the provider's limit and fail the turn, with nothing on screen pointing at the disabled server.
- Too Many Tools No Longer Fails the Turn — When the available tools exceed what a provider accepts in one request, the list is trimmed to fit and the trim is reported once, instead of the request being rejected outright. Nothing withheld is left described in the prompt.
- A Tool That Refused Is No Longer Logged as a Crash — A connector answering something like “unknown device” is an ordinary step the assistant recovers from, not an application fault, and it no longer writes an error and a stack trace into the log you are told to trust. What the model receives is unchanged.
- Failed Requests Explain Themselves — A provider that refuses because the account is out of credit and one that refuses because the model is not in your plan are opposite problems — topping up fixes only the first — and they are now told apart and named. Refusals from providers that report errors in a non-standard shape no longer reach you as a raw error dump.
- Consistent Provider Setup — Preferences and the first-run wizard now show the same provider list, grouped into what is configured and what is not, so the two screens can no longer disagree about what you have set up.
- Wizard Model Lookups — Re-running the setup wizard fired a live model lookup for every provider you had a key for the moment it opened, so an expired key on a provider you were not even looking at produced an error out of nowhere. Models are now fetched for a provider when you open its page.