Deploy and Host Roko

Roko is an orchestration control plane for AI agents that build themselves. It exposes ~85 REST API routes, WebSocket event streaming, and a learning feedback loop. Deploy this template to get a fully functional Roko instance that reads PRDs, generates plans, dispatches agents, validates with gates, and learns from results.

About Hosting

Roko runs as a single container serving HTTP on the configured PORT (default 3000). It persists learning state, episodes, and plans to /workspace/.roko. Attach a volume at that path for durability across redeploys. The health endpoint at GET /api/health returns status, version, uptime, active plans/agents/runs, and provider health.

Environment Variables

• PORT (default: 3000) — HTTP listen port (Railway sets this automatically)
• RUST_LOG (default: info) — log verbosity (debug, info, warn, error)

Common Use Cases

• Orchestrate multi-agent coding workflows with automatic plan generation and gate validation
• Serve as the API backend for the Nunchi dashboard (connect via VITE_ROKO_API_URL)
• Run self-improving AI agent pipelines where cascade routing, prompt experiments, and adaptive gate thresholds persist across runs
• Generate implementation plans from PRDs and execute them with coordinated agent fleets

Dependencies for Roko Hosting

• No external service dependencies — Roko runs fully standalone
• Optionally connect a Mirage instance for on-chain agent coordination
• Optionally connect Roko Workers for distributed task execution

Deployment Dependencies

• Volume mounted at /workspace/.roko for persistent state (learning data, episodes, plans, knowledge store)
• No databases or external caches required — all state is file-based

Why Deploy

• Self-hosting agents: Run your own agent orchestration platform without depending on third-party SaaS
• Dashboard backend: Powers the Nunchi dashboard with real-time agent status, plan progress, and learning analytics
• Learning loop: Agents improve over time — the system learns which models work best for which tasks, tunes gate thresholds, and runs prompt A/B experiments automatically
• Full API surface: ~85 REST endpoints cover agents, plans, PRDs, gates, episodes, signals, knowledge, learning, inference, and more