Commands.com
Platform Docs
A local-first platform for running AI rooms — orchestrated workspaces where independent models converge on optimal solutions through empirical cycles. Your code never leaves your machine.
TL;DR
Run rooms locally where multiple AI models explore, build, and audit independently. Share access through an encrypted relay. Extend with plugin repos. Code never leaves your machine.
On this page
Platform Overview
Multi-model convergence, locally.
Commands.com is a local-first platform for running AI rooms — orchestrated workspaces where multiple models explore, build, and audit independently. Each room follows a convergence pipeline: fan out across explorers, synthesize consensus, build and benchmark, then audit for production readiness.
The relay is available as a hosted service at api.commands.com — no setup required — or you can self-host using the open-source gateway repo.
Use Cases
SQL query optimization — explorers discover index strategies and rewrites, builder benchmarks against real Postgres, auditor validates for production safety
FFT kernel autotuning — explorers search SIMD strategies for Apple Silicon, builder compiles and benchmarks, auditor validates numerical accuracy
Share your codebase — let a colleague ask questions about your code through an AI agent, without giving repo access
Build custom rooms — write a room plugin for any domain-specific convergence workflow and share it with the community
Rooms & Convergence
Fan out. Synthesize. Build. Audit.
A room is an orchestrated workspace where multiple AI models converge on a shared objective through a structured pipeline. Each room plugin defines domain-specific logic for exploration, building, and auditing.
Cast the widest net
Multiple models with different search corpora and training data discover candidates independently. Diversity maximizes signal.
Implement with precision
The strongest agentic coder takes consensus candidates and implements, benchmarks, and iterates until the solution is proven.
Validate for production
An independent model validates operational risk, compliance, and correctness. If the auditor clears it, it's survived genuine scrutiny.
Convergence Pipeline
Fan Out
Multiple explorers discover candidates independently across different models and corpora
Synthesize
Merge, deduplicate, and rank proposals by consensus — promote what independent explorers agree on
Build
Implement top candidates, benchmark with real metrics, iterate until empirically proven
Audit
Independent validation for production risk. Cycle repeats until convergence threshold is met
Why Convergence Works
When independent models with different training data and architectures arrive at the same answer, that's real confidence. The relay only sees encrypted messages — your code, commands, and data are never visible to us.
Getting Started
Up and running in three steps.
Get the desktop app
The agent-workspace repo is currently in private beta. Request access and we'll send download links and onboarding details. Supports macOS, Windows, and Linux.
Launch and connect
Open the app — the setup wizard walks you through agent profile, permissions, and MCP modules. Click Start Agent to open OAuth sign-in and register your device with the relay at api.commands.com.
The workspace includes both a desktop app and CLI — CLI instructions are included with beta access.
Create a room
Install a room plugin, set an objective, configure the room, and launch. Agents are assigned roles and begin the convergence pipeline automatically. Monitor progress through the room dashboard — or share access with collaborators through an encrypted link.
Security Model
End-to-end encrypted. Zero-knowledge by design.
Industry-standard cryptographic protocols ensure your data remains private and secure — architecturally, not just by policy.
Key Exchange
- X25519Elliptic curve Diffie-Hellman for key agreement
- EphemeralFresh keys for every session
- SHA-256Transcript hashing prevents MITM attacks
- HKDFKey derivation with session-specific salt
Message Encryption
- AES-256-GCMAuthenticated encryption with 128-bit tag
- NoncesSequence-based, direction-specific
- AADSession metadata included in authentication
- ReplaySequence numbers detect duplicate messages
Authentication
- OAuth 2.0Firebase Auth with Google identity
- Ed25519Device key signatures for identity
- JWTShort-lived access tokens
- 0600Keys stored locally with restricted permissions
Features
What's included.
Room Dashboard
- Real-time phase progression and cycle tracking
- Candidate frontier table with rankings
- Metrics and improvement tracking
- Winner queries and code output
Desktop App
- Guided setup wizard for agent profiles and MCP modules
- One-click start/stop with live log streaming
- Audit trail viewer with filters and JSON export
- Room creation wizard with plugin-specific config
Permission Profiles
read-only— Read, Grep, Glob onlydev-safe— all tools except blocked; path restrictionsfull— unrestricted access to all tools- You control what remote users can do
Plugin System
- Room plugins for domain-specific workflows
- Agent plugins for model providers
- Interface plugins for external integrations
- Declarative room.yaml schema
Troubleshooting
Common issues.
Device shows as offline
- Ensure the local agent is running:
commands-agent status - Check your network connection and firewall settings
- Verify gateway URL is
https://api.commands.com
Room stuck in a phase
- Check agent logs for errors in the room dashboard
- Verify model provider API keys are configured
- For Docker-based rooms, ensure Docker is running
Agent won't start
- Verify Node.js 20+:
node --version - Build the project:
npm run build - Check config at
~/.commands-agent/config.json
Connection fails during handshake
- Clear browser cache and cookies for Commands.com
- Try signing out and signing back in
- Ensure WebCrypto API is available (HTTPS required)
Still stuck? Contact support@commands.com with your device ID and error logs.
FAQ
Frequently asked questions.
What is a room?
A room is an orchestrated workspace where multiple AI models converge on a shared objective. Each room follows a pipeline: explore, synthesize, build, audit. Room plugins define domain-specific logic for each stage.
Can Commands.com see my data?
No. All messages are encrypted end-to-end using keys only your browser and local agent possess. The relay only sees encrypted data and cannot decrypt it.
Why multiple models instead of one?
One model playing five roles has one set of blind spots. Independent models with different training data and architectures provide genuine diversity. Consensus from independent perspectives is more reliable than agreement from the same model.
Can I build my own room plugins?
Yes. Room plugins are JavaScript modules with a manifest and a room.yaml schema. Define your own convergence workflow for any domain — the platform handles orchestration, dashboards, and lifecycle management.
Do I need to keep my machine on 24/7?
Only when you want rooms running or remote access. The agent must be running for connections. Stop and start it anytime — it reconnects automatically.
Is this open source?
The gateway and room-plugins are open source. The agent-workspace is currently in private beta.
Ready to get started?
Set up your first room and start converging in minutes.
Need help?
Our support team is here for questions, issues, or partnership inquiries.