Product

Changelog

What changed and when. Quiet by design, because we would rather ship a few solid things than a stream of noise.

v0.1.1 · August 2026

Plan mode rebuilt as a vertical pipeline that converges on one answer, instead of handing you a wall of parallel output to sort out yourself.
Vision works without a local Ollama, through Mistral.
The desktop build now reports which engine build it ships, so a bug report says something.
Running one prompt across several models and merging the answers now goes through plan mode, so there is one mechanism behind it instead of two. Verify and ensemble keep working exactly as before.

First signed macOS desktop build, with the engine bundled as a sidecar and an updater that checks for new versions at launch.
Cloud-only operation with your own API keys, so the hosted plans work on a machine with no local models on it.
Clean path-based routing in the app.

What the engine could already do by the time the desktop app shipped:

Orchestration engine: capability map, the <5ms classifier, and capability-scored smart routing.
Models: first-class Ollama with auto-discovery, plus bring-your-own-key for Claude, OpenAI, Mistral, and Groq.
Agents: persistent personas with memory, pilot mode (tool calling), and plan mode (conductor + workers).
Built-in voice: local Whisper speech-to-text, with an optional Mistral Voxtral fallback.
MCP server, and an OpenAI-compatible endpoint at /api/v1/chat/completions.
The local benchmark suite: 16 categories, re-runnable on your own hardware, feeding the capability map that routing reads.