In partnership with 23Blocks ↗ Open source · MIT
Agent orchestration · mission control

Stop juggling terminals. Conduct your agents.

AI Maestro turns scattered coding agents into a coordinated team — persistent memory, live code-graph understanding, agent-to-agent messaging and a decentralized peer mesh, all from one dashboard. Model-agnostic, machine-agnostic, and built so one person can run the output of a whole company.

Visit website ↗ View on GitHub ↗ See what it does
machines, one mesh
Any
agent · any LLM
100+
agents, one operator
~100ms
code-graph re-index
AI Maestro v0.19 localhost:23000 peers
Agents
product
backend-architect claude
frontend-developer cursor
growth
content-writer aider
seo-analyst codex
ops
data-analyst local
backend-architectfrontend-developer
● Reading agent memory (CozoDB)… 1,284 nodes
Implemented GET /api/users — paginated, cached 5m.
→ messaging frontend-developer: "users API ready"
MessagesMemoryGraph
backend-architecthigh
GET /api/users endpoint ready — paginated, supports ?page&limit.
notification · 2m ago
code graph · 5 nodes
One dashboard — hierarchical agents, live terminals, memory, messages and code graph

One person. A whole company's worth of agents.

Product, growth, ops, research — each a team of agents that remember context, coordinate directly and run around the clock across every machine you own. AI Maestro is the mission control that lets a single operator direct all of it, so a tiny team ships like a large one.

Build
write & ship the product
Run
operate the business
Scale
add machines, not headcount
What it does

Four things that turn tools into a team.

Every agent — whatever CLI or model it runs — becomes a coordinated teammate that remembers, understands your code, and talks to the others.

01

One dashboard, unlimited machines

A decentralized peer mesh where every node is equal — no central server. Add a host once and both sides auto-discover; remote agents behave exactly like local ones via transparent WebSocket proxying, with green/red/yellow health at a glance and Tailscale for encrypted links.

Peer meshAuto-discoveryTailscale
MacBook
Mac Mini
AWS
02

Agents that remember

Every agent gets persistent memory in an embedded CozoDB graph. A background "subconscious" indexes conversations for semantic search, consolidates long-term memory, and self-staggers so it scales past 100 agents without CPU spikes. No more re-explaining context each session.

CozoDBSemantic searchSubconscious
indexed conversation #1,284
embedding · all-MiniLM-L6-v2
consolidated 42 → 6 facts
next pass in 4m (idle)
03

Agents that understand code

A live Code Graph, parsed with ts-morph across Ruby, TypeScript, Python and more, maps classes, functions and components and how they import, call and extend each other. Delta indexing keeps it current in ~100ms, so agents know the shape of your codebase before they touch it.

ts-morphMulti-languageDelta index
classfncomponent
04

Agents that talk to each other

A dual-channel messaging system: persistent file-based messages with priorities and types, plus instant tmux notifications for the urgent stuff. Agents coordinate directly — your backend agent hands off to frontend without you playing messenger — and a Slack bridge brings the whole team in.

Inbox / outboxtmux alertsSlack bridge
backend → frontendhigh
users API ready
qa → backendurgent
prod db check
frontend → qanormal
integrated, verify
And more

Universal agent support

Claude Code, Aider, Cursor, Copilot CLI, Codex or your own scripts — anything terminal-based, on any model. Built on tmux + xterm.js with WebGL, streamed over WebSocket.

A zoomable agent canvas

Step back into a spatial “immersive” view of your entire workforce — every agent as a live card you can zoom, arrange and jump into. Mission control at a glance.

Portable agents

Export, import, transfer and clone agents as .zip — metadata, settings, message history and repos — across any machine in the mesh.

Skills, hooks & CLI

A drop-in plugin adds planning, memory-search, docs-search, graph-query and messaging skills, plus 3 hooks and 32 CLI scripts — installed in one command.

Conversation memory

Browse every conversation an agent has had, with thinking and tool usage, searchable semantically across all of them.

Auto-generated docs

Living documentation extracted straight from your codebase and indexed for search — no manual upkeep.

Works from your phone

A fully mobile-optimized dashboard with touch controls — monitor and steer agents from anywhere over Tailscale.

Built with
Next.js 14React 18CozoDBts-morphTransformers.jsxterm.js · WebGLnode-ptyCytoscapeWebSocketFramer MotiontmuxSpace Grotesk
A FluidMind × 23Blocks collaboration

Open-source infrastructure for the agent era.

Free to use, MIT-licensed, made with ♥ in Boulder, Colorado — and shipped alongside the products we run on it every day.

Visit AI Maestro ↗ Star on GitHub ↗ 23blocks.com ↗