Open Multi Agent
Self-hosted TypeScript AI agent framework with MIT licensing, local Ollama support, durable approvals, and run records you can inspect offline.
Open Multi Agent is a self-hosted TypeScript framework for developers building AI agent teams that need human approval and an inspectable record of their work. It runs on your own infrastructure and is free and open source under the MIT license. There's no OMA account, hosted control service, or telemetry.
Teams can use local LLM backends such as Ollama, vLLM, and llama-server, or connect to OpenAI-compatible endpoints and cloud providers including Anthropic, OpenAI, and Gemini. Agents in the same team can use different models. Local or air-gapped models let inference stay in your environment; cloud models send requests to the provider using your credentials.
Approvals can pause a proposed plan, a task, or a consequential tool call. Requests survive restarts and bind the decision to the content the reviewer saw. Tools start with access denied, and network policies can restrict connections made by built-in model adapters.
You can define task dependencies yourself or let a coordinator divide a goal among agents. Independent tasks run in parallel. Checkpoints let interrupted runs resume while reusing completed tool results, and Claude Code, Codex, and Gemini CLI can participate in the task graph.
The optional journal records model context, tool calls, results, and context changes for offline verification. Missing evidence produces an inconclusive result. Governance checks assess which required roles ran and in what order, separately from task success. The offline Run Viewer shows task assignments, tool activity, token usage, and cost estimates.
