Skip to content
AirLink

AirLink

Many humans. Many agents. One full-duplex live workspace.

Section titled “Many humans. Many agents. One full-duplex live workspace.”

The open-source multiplayer room where people and AI agents work, talk and build together in real time.
Live docs, sheets and slides · a shared browser · a Linux desktop · voice agents you can interrupt
Bring any agent: Claude Code, Codex, Gemini CLI or any LLM. Self-hosted. MIT.

Licence: MIT CI npm Node 24

Website · Quickstart · Docs · Why AirLink · Hosted waitlist

AirLink: two people ask two agents for work in chat, then all four edit a sheet, a doc, a flow and a deck at once, then a huddle where the agents answer out loud and edit with their own cursors

▶ Watch with sound

10 ms 42 ms 500
p99 keystroke to every participant agent edit to your screen live clients on one laptop
Section titled “Why AirLink”

Your team works in one place. Your agents work somewhere else: in a chat box, in a remote VM, in a tab you check on later. You prompt, you wait, you copy the answer back into the real work.

AirLink puts agents inside the work. They edit the doc you have open, fill the sheet while you watch, drive the browser you’re both looking at, and talk in the call, where you can interrupt them mid-sentence. People and agents share one live room and one event log, so every participant, human or AI, sees the same state at the same moment.

What you get

Section titled “What you get”

Agents that co-edit, live

Section titled “Agents that co-edit, live”

Rich-text docs, spreadsheets with formulas, slide decks and an Excalidraw canvas. Agents edit with real cursors next to yours. Export to .xlsx and .pptx.

One screen, many drivers

Section titled “One screen, many drivers”

A shared Chromium and a full Linux desktop, streamed over WebRTC. People and agents drive the same screen at the same time.

Full-duplex voice

Section titled “Full-duplex voice”

Start a huddle and agents listen, talk back in their own voice, and stop the moment you cut in. Long tasks run in the background and get reported back to the call.

Bring any agent

Section titled “Bring any agent”

Claude Code, Codex, Gemini CLI or opencode from your PATH, or any model from Anthropic, OpenAI, Google, OpenRouter, Groq, xAI, Mistral or DeepSeek. Plug any harness in over MCP.

Agents with guardrails

Section titled “Agents with guardrails”

Every agent gets its own scoped permissions per object. Prices, dates and commitments arrive as suggestions a person accepts, never as silent edits.

Memory you can read

Section titled “Memory you can read”

Each agent remembers in a folder of plain Markdown files you can open, edit and version, not a black-box vector store.

How it works

Section titled “How it works”
People and agents connect to one live room with one event log; the room holds docs, sheets, slides, a shared browser, a Linux desktop and a voice huddle; agents carry their own memory and guardrails

Every keystroke, click and spoken line, from a person or an agent, lands on one shared event log. Agents don’t watch a copy of your work. They work on the same live objects you do.

Section titled “AirLink vs. computer-use agents”

Screenshot-driven computer-use agents work alone in a VM, one screenshot at a time. AirLink agents work with you, on live structured state.

Computer-use agent AirLink
How the agent sees A screenshot per step Live structured state of every doc and sheet, plus a real-time video stream
Agent action → you see it Seconds per step 42 ms
Working with people Hand off, wait, take over Same doc, same browser, same moment, with cursors
Voice None Full-duplex huddle with barge-in
Memory between steps Rebuilt from pixels Persistent event log of every change, and by whom
Guardrails Whatever the VM can reach Scoped per agent and per object; commitments need a human
Where it runs A hosted VM Your machine, your keys

And when there’s no API, AirLink agents still have the mouse and keyboard on a real Linux desktop.

An AirLink agent drawing a flow on a live canvas with its own cursor while a person watches

What you can build

Section titled “What you can build”
  • An AI teammate in your workspace. Ask in chat; the agent fills the budget sheet, drafts the plan doc and sketches the flow, live in front of you.
  • A voice agent in the room. Talk through a problem in a huddle while the agent listens, answers, and works on the shared screen.
  • A research agent with a shared browser. Watch it browse, point at what matters, and take the wheel whenever you want.
  • Your own product on top. npm create airlink scaffolds a React room, a Node agent, or the full stack, with the client SDK and room tools ready to go.

Quickstart

Section titled “Quickstart”

You need Docker, Node 24 and one LLM key.

Terminal window
npm install
cd apps/airlink
cp .env.example .env            # add ANTHROPIC_API_KEY or OPENAI_API_KEY
docker compose up -d            # Postgres, LiveKit, the shared browser and the desktop
npm run dev                     # → http://localhost:3030

Open http://localhost:3030 in two browser windows, pick a name in each, and say hi in #general. The agent answers, and the dock along the bottom adds a doc, sheet, canvas, slides, a browser, the desktop or a huddle. Add an ElevenLabs or OpenAI key in Settings to hear agents talk.

Building your own app instead:

Terminal window
npm create airlink@latest my-app full     # or room-react, agent-node

Full setup and optional services are in Getting started. AirLink listens on localhost by default; read Security before opening it to a network.

Want it hosted for your team instead? Logins and invites, no Docker, managed desktops and voice. We’ll build it if enough teams ask: join the hosted waitlist.

An AirLink huddle: Lin asks two agents for work out loud, both answer in the call, then they edit the budget and the deck with their own cursors while Lin and Grace write the plan
▶ Watch the huddle with sound

Benchmarks

Section titled “Benchmarks”

Real clients, real Yjs updates, real Chrome. Apple M3, one machine:

p50 p99
Keystroke → every other participant 2 ms 10 ms
Agent edit → a person sees it 42 ms 61 ms
50 rooms × 10 clients, ~2,500 updates/s 1–2 ms 9–48 ms
Room cold start with 10,000 updates 61 ms
Shared browser: scroll → new pixels on screen 127 ms
Run the benchmarks yourself
Terminal window
cd apps/bench
LIVE_JWT_SECRET=<secret> node src/bench.ts --server ws://localhost:3020 --rooms 1 --clients 2 --seconds 30
LIVE_JWT_SECRET=<secret> node src/load.ts --rooms 50 --clients 10 --rate 5 --seconds 60
LIVE_JWT_SECRET=<secret> node src/video.ts --samples 20

Status

Section titled “Status”

AirLink is young: v0.2.0, first public release in October 2026, and much of it was written with Claude Code. What to know before you run it:

  • There are no accounts yet. Anyone who can reach the server can join under any name, so keep it on localhost or put your own auth in front of it. See Security.
  • One Docker desktop per server; set an E2B key for one desktop per object.
  • Voice needs a speech key (ElevenLabs or OpenAI). Without one, agents answer in chat.
  • One voice model, LiveKit’s end-of-turn model, is under the LiveKit Model License, not MIT. See the notice.

We build our own product, MemoAir, on these packages. AirLink itself stays MIT and self-hosted. If you’d want a hosted version for your team, join the waitlist.

Documentation

Section titled “Documentation”

Getting started · Architecture · Agents · Huddle · Desktop · Self-hosting · Security · FAQ

Community and support

Section titled “Community and support”

Contributing

Section titled “Contributing”

Issues and pull requests are welcome. Start with CONTRIBUTING.md and the Code of Conduct; report vulnerabilities through SECURITY.md.

Licence

Section titled “Licence”

MIT.