Self-hosted agent framework where GPT, Claude, and Gemini share a desk

What it does Open-Agent is a self-hostable TypeScript framework pitched as an open alternative to Claude Agent SDK, ChatGPT Agents, and Manus. It bundles multiple frontier LLMs—OpenAI, Claude, Gemini, and open-source models—into a multi-agent system meant to plan, score options, and execute tasks across computers, browsers, and phones while keeping the human in the loop for final calls. The README stays high-level on implementation details, but the promise is a collaborative agent swarm you can modify and host yourself.

The interesting bit Instead of anointing a single model as the brain, it orchestrates several frontier LLMs to collaborate on the same task, using spec and context engineering to structure planning and surface options for human approval. It is essentially a bet that a committee of competing APIs is more reliable than one proprietary black box.

Key highlights

• Self-hostable stack, Apache 2.0 licensed.
• Orchestrates OpenAI, Claude, Gemini, and open-source models in a single multi-agent workflow.
• Planning layer uses spec and context engineering to score options and defer final decisions to the operator.
• Acknowledges architectural inspiration from AFFiNE and the broader open-source agentic community.
• Live demo available at open-agent.io.

Verdict Worth a spin if you want a vendor-agnostic, self-hosted agent foundation you can reshape. Look elsewhere if you need exhaustive docs or proven benchmarks today—the README stays high-level and leaves the plumbing as an exercise for the reader.