docs

Introduction

Ninetrix is an open-source platform for building, shipping, and observing AI agents as Docker containers. agentfile.yaml is to AI agents what Dockerfile is to containers — a portable, version-controlled definition that works anywhere.

Open source, Apache 2.0
Ninetrix is fully open source. Self-host the entire stack on your own infrastructure, or use Ninetrix Cloud for a managed experience.

The problem it solves

Building AI agents today means gluing together frameworks, writing boilerplate, reinventing observability, and figuring out deployment on your own. Ninetrix gives you a standard:

  • One YAML file defines your agent — provider, model, tools, memory, triggers, governance
  • ninetrix build compiles it into a Docker image with a full generated runtime
  • ninetrix run executes it with traces, checkpoints, and approvals built in
  • ninetrix up starts a multi-agent network on a shared Docker bridge
  • Self-host the entire stack — dashboard, API, MCP gateway — with one docker compose up

Key features

FeatureDescription
Multi-providerAnthropic, OpenAI, Google, DeepSeek, Mistral, Groq, Together AI, OpenRouter, Cerebras, Fireworks, AWS Bedrock, Azure OpenAI, MiniMax — switch without code changes
MCP-native toolsConnect any MCP server: GitHub, Slack, filesystem, browser, and more
Persistent memoryAPI-based checkpoints — resume any session with --thread-id, no database config required
Durable RunsAutomatic crash recovery — Docker restarts the container, agent resumes from last checkpoint
Multi-agentAgents hand off to each other via collaborators on a Docker bridge
Human-in-the-loopGate specific tool calls on human approval before executing
Planned executionAgent generates a plan first, executes step-by-step with verification
TriggersWebhook and cron schedule triggers built into the container
GovernancePer-run budget limits, approval gates, rate limits
ObservabilityFull trace view in the dashboard — every tool call, every step
Self-hostableOne docker compose up — air-gap ready, no vendor calls

How it works

ninetrix build reads your agentfile.yaml, renders Jinja2 templates into a Python runtime and Dockerfile, then runs docker build. The resulting image contains everything your agent needs — the LLM client, tool connectors, checkpoint logic, trigger servers, and approval handlers.

Bash
# Install (auto-detects pipx, uv, or pip3)
curl -fsSL https://install.ninetrix.io | sh

# Or manually:
# pip install ninetrix
# uv tool install ninetrix

# Start local stack (Postgres, API, MCP gateway, dashboard)
ninetrix dev

# Build and run your first agent
ninetrix build
ninetrix run

Next steps

  1. Quick Start — build your first agent in 5 minutes
  2. Installation — full install guide with prerequisites
  3. agentfile.yaml — the full YAML reference
  4. MCP Tools — connect GitHub, Slack, and any MCP server
On this page