Control Tower v0.2.0
/

Control Tower documentation

Control Tower is a self-hosted AI gateway with a live map of every agentic data flow. Agents call models, MCP tool servers and HTTP APIs through it; you see every path on the map, put gates on the ones that matter, approve risky calls as they happen, and account for every call.

The console: see every flow, trace an agent, draw a gate, approve the held call

Start here

  1. Getting started — from docker run to your own agent on the map, in five minutes.
  2. Install — Docker, Compose, Render, Fly.io, Railway, any container platform, or from source; upgrades and backups.
  3. Connect your agents — OpenAI SDKs, the OpenAI Agents SDK and Codex, Claude Code, LangChain, MCP clients, plain HTTP. Step by step: Claude Code, Claude Desktop, Codex in the ChatGPT desktop app, Codex CLI.

Set up

  • Providers and models — connect OpenAI, Anthropic, Gemini, Bedrock, Vertex AI, Azure, Ollama…; models added on first use; aliases, load balancing and fallbacks.
  • Images, audio and providers' own APIs — images, speech, transcription, moderation, rerank and completions; Gemini's and Bedrock's own SDKs through the gateway.
  • Keys, budgets and limits — one key per agent, with allowed models and tools, rate limits and budgets.
  • MCP tool servers — register tool servers; per-key tool visibility.
  • HTTP APIs and observed traffic — route REST APIs through the gateway, and map what doesn't go through it.
  • A2A agents — put remote agents that speak the Agent2Agent protocol behind the gateway.
  • Agents calling agents — follow calls from one agent to the next, and gate on whom a call is made for.

Control

  • The Airspace — reading the map, gates (block, require approval, inspect), simulation, approvals in the Tower, zones.
  • Guardrail services — Presidio, Lakera, Bedrock Guardrails, Azure AI Content Safety, OpenAI moderation or your own URL, asked by inspect gates.
  • Policy as code — zones and gates as YAML, with a preview before anything changes.
  • Alerts — console, Slack and webhooks; approving from Slack.
  • Monitoring — Flights, the Ledger, the data-flow inventory, Prometheus metrics.
  • Exporting flights — every call to your OpenTelemetry, Datadog, Splunk, S3 or webhooks, as metadata.
  • People and roles — admins, approvers and viewers; adding people; lost passwords.
  • Running several instances — Postgres and Redis behind a load balancer.
  • What is enforced — the boundary, stated honestly.

Reference

  • Configuration — command-line flags, environment variables, the admin key.
  • Config file — config.yaml for providers, models, fallbacks, MCP servers and alerting, field by field.
  • Demo mode — a synthetic fleet to explore with.
  • Architecture — how a request flows, what is stored, retention, limits.
  • API reference — every gateway, admin and health endpoint.
  • Troubleshooting — errors agents see and how to fix them.
  • Changelog — what changed in each release.

Screenshots in these pages are taken from a real server by pnpm build && pnpm docs:screenshots.

Edit this page