23blocks-OS/ai-maestro

AI Maestro – an operating‑system‑style platform for AI‑first teams

What it is – AI Maestro is a self‑hosted dashboard and runtime that lets you run many AI agents (Claude Code, Codex, Aider, etc.) on one or many machines and treat each agent like an employee. Agents have persistent memory, an inbox for messages, a visual avatar, and can own code repositories. The platform coordinates them via the open Agent Messaging Protocol (AMP) and other open standards (AID, AAP).

Key capabilities

Feature	What you get
One unified UI	See every agent, start new ones, organise them in a tree, and switch contexts with a click.
Multi‑machine mesh	Peer‑to‑peer network; any machine (Mac, Linux, Windows‑WSL2, Docker, AWS EC2/Fargate) appears in the same dashboard – no central server needed.
Agent‑to‑agent messaging	Email‑like messages with priority, signatures and push notifications; agents can request work from each other automatically.
Persistent long‑term memory	Nightly summarisation of an agent’s conversations into a searchable knowledge graph and a code‑graph that survives restarts.
Work coordination tools	Kanban board, split‑pane “war rooms”, scheduled jobs that travel with the agent, and avatars/TTS for a more human‑like interaction.
Pluggable deployment	Run agents locally in tmux, in Docker, on a dedicated AWS EC2 instance, or as AWS Fargate serverless containers.
Gateways	Connect agents to Slack, Discord, Email, WhatsApp with content‑security filters.
Extensible plugins	Add new skills, CLI tools, or custom actions via the provided plugin SDK.

How it works – An agent is a process that runs a terminal‑based LLM tool. AI Maestro stores the agent’s identity (cryptographic keys), its memory graph, and its code checkout. When you create an agent, the UI launches the chosen runtime (tmux, Docker, etc.) and registers the agent in the mesh. The mesh propagates presence information, so the dashboard can list agents on any host. Agents talk to each other through AMP, which handles routing, signatures and optional encryption.

Typical workflow

1. Install – one‑line script (curl … | sh) pulls the dashboard, AMP service and optional Claude Code plugin.
2. Create your first agent – the wizard spins up Lola, a “Chief of Staff” bot that triages email, builds a memory graph and can be extended.
3. Add more machines – install the headless runtime on a spare Mac Mini or an EC2 instance; it joins the mesh automatically.
4. Build a team – give each agent a name, avatar, and a set of skills (e.g., code review, data‑query). Agents can now send messages like @AIM:backend-agent Deploy new version.
5. Monitor & iterate – use the Kanban board, code‑graph visualiser, or playback view to see what each agent has done and adjust prompts or plugins.

Who it’s for

• Founders or solo operators who want a “virtual staff” of AI agents.
• Developers who already run several terminal‑based LLM assistants and need a single place to manage them.
• Teams that need agents on different OSes or cloud resources but want them to cooperate without a custom glue layer.

What you need

• Node .js 18+, tmux, and (optionally) Docker.
• For Windows, a WSL2 Linux distro.
• Any terminal‑based AI tool (Claude Code is recommended for the inbox mod, but not required).

Open‑source stance – MIT‑licensed, no SaaS lock‑in, no telemetry. Core protocols (AMP, AID, AAP) are open standards, and the repo includes plugin templates and a separate LolaBot ecosystem for building custom agents.

Roadmap highlights

• One‑click agent creation from a Git repo URL.
• Global search & filtering across the whole mesh.
• Playback / time‑travel of agent sessions.
• Analytics dashboard for performance metrics.

Bottom line – AI Maestro is not a simple script or a tutorial list; it is a full‑featured, self‑hosted operating‑system‑like layer that lets you run, persist, and orchestrate many AI agents across machines, turning a collection of LLM assistants into a coordinated “AI‑first organization.”