Open-source managed agent runtimes

Open-Source Managed Agent Runtimes Compared

A managed agent runtime owns agents, sessions, sandboxes, and crash recovery behind an API. OpenMA, TrueForge, and Omnara are open-source runtimes with that shape, and Claude Managed Agents is the hosted reference.

The managed runtime shape

A managed agent runtime is a service with an HTTP API. You create an agent and an environment, open a session, send a message, and stream events back while the runtime runs the loop, executes tools in a sandbox, persists every step, and recovers from failures. Open-sourcing that layer means you can audit how credentials and data are handled, run it where your data must live, and change the parts that do not fit.

For the platform primitives behind these runtimes, see the AI agent platform guide. For the hosted cost baseline, see Claude Managed Agents pricing.

Criteria for the managed-runtime shape

Use these as a checklist when comparing runtimes. A project missing several of them is a framework or harness, which can still be the right tool for your use case.

Managed agent runtime checklist
Capability	What to check
Agent resource	Versioned agent definitions: model, prompt, tools, MCP, skills
Environment resource	Declarative sandbox spec: packages, networking
Durable sessions	Long-running, resumable, recover after crashes
Event stream	Append-only event log streamed over SSE
Isolated execution	Tools in a sandbox, not in the API process
Vaults	Credentials kept in a vault and added to outbound requests, so the sandbox does not hold them
Self-hostable control plane	The whole runtime runs on your infrastructure

Open-source managed agent runtimes compared

Claude Managed Agents is included as the reference shape. Third-party details come from each project's repository (TrueForge, Omnara) as of November 2026.

Open-source managed agent runtimes shortlist
Runtime	Managed-agents API	Self-host control plane	Models	License
OpenMA	Yes: Claude Managed Agents-compatible /v1 plus OpenAI Agents API adapter	Yes: Node/Docker, Fly.io, Vercel (beta), Cloudflare	Claude, OpenAI, DeepSeek, OpenRouter, custom	Apache 2.0
TrueForge	Own HTTP API and TypeScript SDK	Yes: local SQLite, or Postgres + Redis	OpenAI, Anthropic, Gemini, OpenAI-compatible	MIT
Omnara	Own REST API	Yes: Docker Compose with Postgres, Redis, S3	Any compatible model	Apache 2.0
Claude Managed Agents (reference)	Defines the shape	No — self-hosted sandbox only	Claude	Proprietary

For the wider field of frameworks, harnesses, and coding agents, see open-source AI agents.

Self-host next steps

Start OpenMA locally with Docker Compose, then move to a Node deployment with PostgreSQL or MySQL, or provision it on Cloudflare. The self-host overview lists every path, and self-hosted AI agents explains the trade-offs.

vs Claude Managed Agents

If you are choosing between Anthropic's hosted runtime and an open one, read the Claude Managed Agents alternative guide and the feature-level Claude Managed Agents vs OpenMA comparison.

FAQ

What is an open-source managed agent runtime?

A runtime that implements the managed-agents shape (agents, environments, sessions, events, sandboxes, and credential isolation) with source you can self-host and change.

Which open-source runtime implements the Claude Managed Agents shape?

Claude Managed Agents itself is proprietary. OpenMA is an Apache 2.0 runtime that implements the same agents, environments, sessions, and events model and can be self-hosted.