Roadmap

What's shipped, what's being built, and what's planned.

Shipped

Version	Feature
v0.1	Local GGUF model support via llama.cpp — fully offline AI, no cloud required
v0.2	Bundled skills system (Cerebellum) — 49 knowledge modules the agent loads pre-run
v0.3	Persistent memory (Hippocampus) — SQLite-backed context retention across sessions
v0.4	Desktop automation — screen capture, click, type, OCR, focus windows
v0.5	Basal Ganglia self-improvement — auto-creates skills from repeated tool patterns
v0.6	Sandbox execution — isolated environments with diff/patch/promote workflow
v0.7	Scheduler — recurring agent jobs, cron-style, with full run history
v0.7	208 tools across files, shell, git, browser, AI, memory, Docker, audio, artifacts
v0.7	20+ AI providers — local (llama.cpp, Ollama, MLX) + cloud (OpenAI, Anthropic, Gemini, …)
v0.7	Integrations — GitHub, Google Calendar, Outlook, Obsidian, email, RSS
v0.7	Multi-agent delegation — route work to other agents or profiles mid-session

In progress

Item	Notes
MCP protocol support	Full Model Context Protocol client — connect to the broader MCP ecosystem
Team server mode	Multi-user support with role-based access control (`SOLO_MODE=false`)
API / SDK	Programmatic access to agent runs and workspace data

Planned

Item	Notes
Multi-platform clients	Telegram, Slack, and web portal clients that connect to the same backend
Self-improving skill loop	Basal Ganglia generating and benchmarking skills autonomously over time
Plugin marketplace	Shareable community skills and tool bundles
Mobile companion	iOS/Android app for accessing asyncat remotely

Contributing

asyncat is open source under MIT. Issues, PRs, and feature requests are welcome on GitHub.

The codebase is organized around three packages:

den — Express backend, agent runtime, all tools and skills
neko — React + Vite frontend
electron — Electron desktop container wrapper

See Development for how to run from source.