AgentMesh — Multi-Agent Orchestration Framework

Overview

Single-prompt AI hits a ceiling on complex work — real tasks need research, then analysis, then drafting, then review, each done well. AgentMesh is the framework for orchestrating that. You register specialist agents (each just a name, a role, and a system prompt), and a coordinator — running your strongest model — reads the task, produces a plan of which agent does what, delegates each step (passing along the work done so far), and synthesizes everything into one polished result. Crucially, it's transparent: every run returns the plan, each agent's individual output, and the final answer, so you can debug and tune instead of fighting a black box. It ships with four working example agents (researcher, analyst, writer, critic) that you replace with your own — legal reviewers, translators, data extractors, whatever your domain needs. This is base code to build products on, not a single-purpose app. It runs on your own servers with your own Anthropic API key.

Key features

• Coordinator automatically plans and delegates across your agents
• Define agents in one file — name, role, system prompt (and optional model)
• Full run trace: the plan, every agent's output, and the final answer
• Per-agent model choice — a strong coordinator, cheaper workers
• Four working example agents included, ready to replace with your own
• Transparent and debuggable — never a black box
• Bring your own Anthropic API key — no per-seat limits

Specifications

Product type

Self-hosted software framework (source code)

License

Perpetual, modify & redistribute in your products

Backend

Python 3.12 · FastAPI

Frontend

Next.js · React (run console)

Coordinator model

Claude Opus (configurable)

Worker model

Claude Sonnet (configurable, per-agent)

Requirements

Docker + your own Anthropic API key

Delivery

License key + private source download link