Local AI rooms where independent models converge.
Orchestrated rooms running on your machine. Multiple AI models explore, build, and audit independently — converging on solutions proven by real metrics, not model confidence.
One model playing five roles still has one set of blind spots. Giving it five personas doesn't give you five perspectives — it gives you one perspective in five costumes.
Commands.com runs rooms — orchestrated workspaces where independent models with different training data, architectures, and tooling converge on a shared objective through a structured pipeline.
The most reliable AI is not one model. It's independent models that converge.
Every room follows a convergence cycle:
Fan Out → Synthesize → Build → Audit
Multiple Merge, dedup, Implement top Independent
explorers rank by candidates, validation for
discover consensus benchmark with production risk
independently real metrics
Explored: 12 → Synthesized: 5 → Built: 3 → Converged: 1
| Role | What it does | Why a separate model |
|---|---|---|
| Explorer | Casts the widest net across candidates | Different search corpora and training data maximize diversity of discovery |
| Builder | Implements with precision, benchmarks empirically | The strongest agentic coder, informed by consensus from explorers |
| Auditor | Validates for production safety and correctness | Independent scrutiny catches what the builder's biases miss |
| Repo | What it does | Stack |
|---|---|---|
| gateway | Self-hosted encrypted relay, OAuth/OIDC, device routing, shares, webhook tunnel, admin console | Go |
| agent-workspace | Local runtime, rooms engine, desktop app, CLI, audit logging, remote chat, MCP passthrough | TS / Electron / Node |
| room-plugins | Room orchestrators for domain-specific convergence — SQL optimization, FFT autotuning, and more | JS plugins |
| interface-plugins | External interface providers for webhook and Slack-style integrations | JS plugins |
| agent-plugins | Provider plugins that make model diversity real instead of cosmetic | JS plugins |
Postgres Query Optimizer — Explorers independently discover index strategies and query rewrites. The builder implements each candidate against a real Postgres instance, measuring actual execution time. The auditor validates for production safety.
FFT Autotune — Explorers search for SIMD vectorization strategies across Apple Silicon architectures. The builder compiles and benchmarks each kernel variant. The auditor validates numerical accuracy. Convergence means the fastest correct implementation wins.
- End-to-end encrypted relay (AES-256-GCM) — we can't read your data, even if we wanted to
- Rooms, agents, code, and data stay on your machine
- Self-hosted gateway option for teams that want infrastructure ownership
- Zero-knowledge architecture — by design, not by policy
Sign up for the beta to get early access to the desktop app and start running rooms locally.
# Self-host the relay
git clone https://github.com/Commands-com/gateway.git
# Extend with plugins
git clone https://github.com/Commands-com/room-plugins.git
git clone https://github.com/Commands-com/agent-plugins.git
git clone https://github.com/Commands-com/interface-plugins.git