Exo Agent Framework

Exo Agent Framework is a self-hosted framework for developers building long-running AI agents that can revise their own code, tools and prompts. Its exoharness foundation keeps agent history and resources separate from the logic that decides what the agent does next. The project is open source under the MIT license.

That separation is its main distinction. An agent can change its implementation across runs while exoharness retains its identity, event history, secrets and artifacts. Developers can stop and resume a run, replace the executor with Codex, Claude Code, Cursor or their own implementation, and fork or rewind from an earlier event without losing resources. The recorded history remains the source of truth even when the agent's prompt or execution logic changes.

The framework has a Rust core and TypeScript support for custom agent harnesses. It works with local models as well as OpenAI, Claude and Gemini APIs. The harness runs on your own machine or server; choosing a cloud model sends model requests to that provider.

For agents that run commands and edit files, sandboxes provide workspace access, network controls and snapshots across turns. Exo also supports command-line chat and browser access through ExoChat, plus messaging connections such as Discord, Slack and Signal. Agents can use scheduling and human-authored extensions for task-specific work.

It's early software, with an unstable public API. Exo can inspect its own code and runtime logs, clone itself and manage a lineage of clones.