Warning
Mandri is in alpha - expect bugs and breaking changes.
Mandri is an open, local-first orchestrator for autonomous agents.
For a simple task, Mandri can delegate the work to a single agent. If that agent gets stuck, it can bring in another agent with a different model, runtime, or set of capabilities.
For complex tasks, Mandri coordinates multiple specialist agents, allowing them to exchange requests, findings, and results while working toward the same outcome.
Mandri does not try to rebuild coding, browser, or terminal agents from scratch.
It orchestrates agents that already know how to do their job, such as Claude Code, Codex, OpenCode, and other specialized runtimes, and gives them a way to delegate work, exchange context, and coordinate around the same task.
Each agent keeps its own tools, execution loop, context, permissions, and strengths.
Mandri separates the agent runtime from the language model behind it.
Through its model gateway, a compatible agent can use the provider or model selected by the user, whether it runs locally, through a cloud API, or behind any OpenAI-compatible endpoint.
Mandri uses as few dependencies as possible.
It is not an all-in-one agent framework, a workflow platform, or a replacement for every tool it connects to. Its purpose is narrower: delegate work, coordinate agents, route communication, and keep the user in control.
Mandri is licensed under the Apache License 2.0.