The open-source, self-hosted foundation for AI agents.
Agent Foundation is an open-source library and platform for building and running your own agent systems. Managed agents, memory, sandboxes, computer use, and durable execution come together in a self-hosted service, ready to integrate into your applications.
- Build with Service. Configure agents and connect them to your product through APIs, with resource management, permissions, and execution recovery already in place.
- Extend with Harness. Embed the runtime directly and shape its behavior through plugins, custom tools, and providers.
- Explore with Harness UI. Try models, tools, and agent configurations in a terminal and web playground—without building an application first.
These guides track main. For an installed release, use its release notes and package metadata.
Run Service with Docker Compose
You need Docker with Docker Compose and either a model provider API key or an eligible ChatGPT plan. The stack starts Service, Console, PostgreSQL, and Redis. Agents use your host Docker Engine for execution environments.
Console — Build agents, try runs, inspect execution traces, and monitor usage. Shown with fictional demo data.
Download the Compose file and start the stack:
mkdir a13n-service
cd a13n-service
curl -fL https://raw.githubusercontent.com/converge-ai-labs/agent-foundation/main/deploy/docker/compose/a13n-service.yaml -o a13n-service.yaml
docker compose -f a13n-service.yaml up -d --wait --pull always
Then open http://127.0.0.1:8080:
- Register the administrator account. Enter your email and a password of at least 8 characters. Service creates your organization and workspace and signs you in.
- Connect a model. Under Models, add a provider using an API key or Sign in with ChatGPT, then select a model.
- Try an agent. Create an agent, choose the model, and send your first message with Try agent.
Sign in with the same account on later visits. Administrators invite additional users.
Already in a repository checkout? Run make compose-up from the repository root to pull images, start the same stack, and print its Console URL.
The stack binds to loopback and mounts the host Docker socket. For a shared deployment, follow the deployment guide to configure HTTPS, bootstrap the administrator, and select a release version.
Continue with the Service quickstart for model setup, API usage, stop/resume, and troubleshooting.
Turn execution evidence into improvements
Findings helps you inspect execution evidence, review diagnoses and record corrections. Hand the cited Agent revision and evidence to Composer for a proposed change, then review and approve it. Nothing is automatically repaired or published. See Find and improve execution issues.
Scan diagnoses and review status. Shown with fictional local seed data.
Inspect the evidence, record your judgment, then ask Composer to propose a change to the cited version.
Choose your agent stack
Build your agent stack in Console: choose model providers, web data services, execution environments, and remote MCP servers.
| Capability | Explore |
|---|
| Models | Model providers, including cloud APIs, gateways, and local Ollama endpoints. |
| Web Data | Search and scrape providers, including Brave, Exa, Tavily, and TinyFish. |
| Environments | Docker and hosted sandboxes, including E2B, Daytona, Modal, Runloop, Sprites, and Vercel. |
| Connections | Browse remote MCP servers or connect your own compatible endpoint. |
Environment execution has two paths: Native Local/Docker/E2B for quick setup with Files & Shell, and Envd for general execution with OS sandbox, Egress, Computer use, and Device metrics. Each infrastructure vendor retains its resource management provider.
Explore the full-size screenshots in the linked guides. The Local directory environment is available only in development.
Sign in with ChatGPT
Harness UI and Service support Sign in with ChatGPT: use your eligible ChatGPT plan for agent requests without an OpenAI API key. See Harness UI setup or Service setup.
Sign in with ChatGPT — Authorize Agent Foundation OSS on OpenAI's sign-in page. Account details are obscured.
Use Harness UI
For individual work or trusted collaborators, Harness UI offers a terminal agent and browser workbench over the same Harness. Explore repositories, edit files, run commands, and try models, tools, Skills, and environments interactively.
Harness UI — Review agent output alongside project files in WebUI. Shown with a fictional project and a local demo model.
uv tool install a13n-harness-ui
cd your-repository
a13n-harness-ui # TUI
# Or: a13n-harness-ui webui
First-use setup connects a model and selects execution permissions. Full Control runs as your host account. Share WebUI only with trusted collaborators; --no-share-computer disables its native host file and terminal access. See installation, execution permissions, and browser access.
Build on Harness
Harness gives your application reusable agents, typed tools and outputs, streaming observations, portable execution environments, and state you can save and resume. Your application owns persistence, credentials, and recovery policy.
We evaluated A13N Harness on Terminal-Bench 4.0 against official Codex using the same model and reasoning effort. Click the chart for the full results and methodology.
Using the FrontierHarness Eval task set and methodology, A13N Harness with Kimi K3 passed 21 of 30 tasks (70.0%) at $2.93 standardized cost per pass. We used our general-purpose instructions, without benchmark-specific instruction tuning to reduce token usage or inference cost. The comparison with published baselines is provisional: local execution conditions differ, and no matched control was run. Click the chart for the full results, accounting, and methodology.
Start with the Harness quickstart, then explore runnable examples. The package catalog covers supporting components such as Envd, Stream Protocol, and logging.
Work from source
git clone https://github.com/converge-ai-labs/agent-foundation.git
cd agent-foundation
Source development uses Python 3.13, uv, and Make. Browser builds also need Node.js 24 and the pinned pnpm version; Service development needs Docker. Follow CONTRIBUTING.md for the toolchain and checks relevant to your change.
Community
Join the a13n community on Slack to ask questions, share what you're building, and discuss ideas with other users and contributors.
Contribute
Bug reports, documentation fixes, examples, and focused improvements are welcome. Search GitHub Issues and read the contribution guide before starting. Discuss unresolved product or architecture decisions before implementing them.
Security
Report suspected vulnerabilities privately to [email protected], not in public issues or pull requests. See SECURITY.md for what to include.
License
Agent Foundation is licensed under the Apache License 2.0.
Copyright 2026 Converge AI.
