Its own agent engine, harness, and prompts — not a wrapper around a terminal. Everything runs locally, and it's yours.
Describe a goal. It fans out into up to seven parallel agents, each with an automatic verify → repair loop, and gives a plain-language "what changed, what's next" summary as each finishes. Model-aware concurrency; optional manual queue.
Cross-project memory learns your stack and conventions once, global search spans every past session, and episodic memory reuses what worked while avoiding past mistakes — plus 100+ bundled skills you invoke with a slash, and your own.
34 built-in providers — Anthropic, OpenAI, Google, Groq, Mistral, DeepSeek, OpenRouter and more — plus local and LAN models via Ollama or LM Studio. Searchable picker, favorites, reasoning variants.
A dedicated review panel stays fast across hundreds of files, click-to-comment diffs collect into the chat, and commit → push → open-PR happens without leaving the app.
An interactive map of how a project's symbols reference each other — Force, 3D, Ring, and Tree layouts.
Schedule recurring agent runs in the background; they land in a human-in-the-loop review state, with auto-archiving and retry-with-backoff.
A mic for voice input and a live avatar that reacts to the session — narrate, mute, collapse to a bar, or pop it out as its own floating window.
SSE streaming, inline tool-call visualization, undo/redo of an agent turn, @ file mentions, a command palette, system tray, auto-updates, and encrypted credential storage.
Each comes with a guided path, real one-click MCP connectors — Figma, GitHub, Supabase, Postgres, Firebase — and vetted example repos.
34 built-in providers — plus local and LAN models. Run a model on your own machine, or point at another box on your network. Searchable picker, favorites, and reasoning variants.
Alpha, under active development. Releases live on GitHub.