Exo Agent Framework

Exo Agent Framework is for developers building long-running AI agents that can revise their own code, tools and prompts. It preserves agent identity and history as their behavior changes across runs.

• Self-modifying agents: Exo can inspect its code and runtime logs, change its implementation, clone itself and manage a lineage of clones.
• Persistent history: Stop and resume an agent, or fork and rewind from an earlier event while retaining its resources and recorded actions.
• Replaceable executors: Choose Codex, Claude Code, Cursor or a custom executor to direct the agent without replacing its stored identity and history.
• Sandboxed work: Agents can run commands and edit mounted workspaces in sandboxes with network controls and snapshots across turns.
• Chat and scheduled tasks: Talk to agents through the command line or ExoChat in a browser, connect Discord, Slack or Signal, and schedule work.

The framework runs on your own machine or server and is open source under the MIT license. It supports local models and OpenAI, Claude and Gemini APIs. Requests that use cloud models go to the selected provider. Developers can build custom agents with its Rust core and TypeScript harness support.

Exo is early software with an unstable public API. It doesn't provide a simple way to perform general computer use in a windowed desktop environment.