What are open-source AI agents?
Agent software whose source is published under an open license. The term spans complete agent platforms, harnesses, frameworks, and coding agents, so compare projects within the same category.
Open-source AI agents
Open-Source AI Agents & Agent Platforms
"Open-source AI agents" lumps together very different projects: complete agent platforms, harnesses, frameworks, and coding agents. Only the platforms run agents behind an API with sessions, sandboxes, and credentials, which is what production use requires.
What "open-source AI agent platform" means
A platform runs agents for you behind an API. A library helps you write agent logic and leaves the running to you. The AI agent platform & orchestration guide compares platforms, SDKs, builders, and coding agents in one table.
Selection criteria
Sessions, sandboxes, multi-model, self-host
| Criterion | What good looks like |
|---|---|
| Sessions | Durable, resumable, survive process crashes |
| Sandboxes | Tools run isolated from the API process, and the provider is replaceable |
| Multi-model | Each agent sets its model in configuration, with no runtime changes |
| Self-host | The whole control plane runs on your infrastructure without a license gate |
| Credentials | Secrets stay in a vault and are added to outbound requests, so the sandbox does not hold them |
| API | A stable HTTP API with streaming, so any language can be a client |
If you are building directly on a model SDK, the Claude Agent SDK comparison shows which of these you would need to add yourself. For the category view, see AI agent platform & orchestration.
Comparison table
Projects are grouped by category in no particular order. Third-party details come from each project's repository as of November 2026.
| Project | Category | Managed-runtime shape | Notes |
|---|---|---|---|
| OpenMA | Orchestration / managed runtime | Yes: agents, environments, sessions, events, vaults, memory stores | Apache 2.0, multi-model, with Claude Managed Agents and OpenAI Agents API compatible endpoints |
| TrueForge | Agent harness | Yes: sessions, HTTP API and TypeScript SDK, approvals, Daytona sandboxes | MIT. TrueFoundry markets it as a CMA alternative |
| AX (Agent Executor) | Orchestration runtime on Kubernetes | Task, Workspace, and Model manifests (ax.io/v1alpha1) | Apache 2.0. Google's declarative orchestrator, runs on Agent Substrate, pre-1.0 |
| Omnara | Managed agent platform | Yes: REST API, durable state, schedules, sandboxes | Apache 2.0, self-hostable with Docker Compose |
| OpenHands | Coding agent (labeled) | No, software-engineering agent | Open-source coding agent, comparable to Claude Code and Codex |
| LangGraph / CrewAI / AutoGen | Frameworks | No, you assemble it | Mature libraries for graphs, roles, and multi-agent conversations |
OpenMA (orchestration / managed runtime)
A meta-harness: it manages sessions, sandboxes, tools, credentials, and recovery around whichever harness and model you choose, including Claude Code or Codex through ACP.
TrueForge, AX, Omnara
TrueForge is an MIT harness with its own API and chat UI. AX is Google's Kubernetes orchestrator that runs agent tasks as sandboxed actors on Agent Substrate, with an API marked v1alpha1. Omnara is an Apache-2.0 managed platform that can run agents on your own machines or on third-party sandboxes. Sources: TrueForge, AX, Omnara.
OpenHands (coding agent)
A strong open-source software-engineering agent. Its peers are Claude Code and Codex. Like them, it can run inside an agent runtime.
Open-source managed agents path
If you want the "managed agents" shape popularized by Claude Managed Agents and the OpenAI Agents API in an open, self-hostable form, read the open-source managed agents shortlist, then self-hosted AI agents for deployment.
Comparing with Anthropic's hosted agents
Evaluating Anthropic's hosted runtime against open options? The Claude Managed Agents alternative guide has the decision table.
Pick a managed-runtime shape, then self-host OpenMA
FAQ
What is the best open-source AI agent platform?
The one that meets the runtime criteria you need: durable sessions, isolated sandboxes, multi-model support, credential isolation, and self-hosting without a license gate. OpenMA covers the criteria in the table above. Frameworks such as LangGraph or CrewAI are libraries you would build a platform around.
Is LangGraph an agent platform?
LangGraph is a framework for agent logic. A hosted runtime is offered separately by LangChain. On its own, it does not provide sandboxes, vaults, or an operator console.
Are open-source agents free?
The software is free to use under its license. You still pay for model usage, compute, and the time to operate it.
Related guides
Claude Managed Agents alternative
Buyers comparing Anthropic-hosted agents with open runtimes.
Open-source managed agents
Shortlist of open runtimes with the managed-agent shape.
Self-hosted AI agents
Run agents on your own infrastructure.
AI agent platform & orchestration
The primitives behind an agent platform.
Claude Agent SDK
SDK vs managed runtime, and when to migrate.