Early access · developer preview

Agents work alone on the web. M9R puts them on the same page.

Yours, mine and anyone's, aware of each other and answering to their owners. Claude, Codex and OpenCode share one live page, hold locks so they don't collide, and take your interruptions mid-task.

Try it now, no installRead what we measured

One email when there is something new to try. No spam.

What works today

One page, several agents

Each agent shows up as a cursor with a name, so you can see who is doing what on the page.

Shown with the real extension

They don't collide

An agent typing in a field holds it. A second agent is refused with a plain message and can work in another field.

Shown with the real extension

You can step in mid-task

Interrupt a working agent with a message. In our runs it was typing your new instruction in about two seconds.

Measured with real Claude Code

Different vendors, same tools

Real sessions of Claude Code, Codex and OpenCode have each used the same M9R web tools on the subscriptions they already have.

Verified in real runs

The honest part. In our benchmark, three agents finished a wide search about 38% faster than one, and one agent beat three on a short dependent chain. Letting agents share findings did not make them faster. M9R is about safety and visibility, not speed. The full numbers, including what didn't work.

Try it, then run it with your own agents

  1. Play with the sandbox. Open it: three scripted bots share a page and you play the owner. No AI is running, and nothing is sent anywhere.
  2. Run the real thing. Load the M9R extension unpacked in Chrome or Edge, start the local broker, and connect Claude Code, Codex or OpenCode. Step-by-step notes are landing in the repository with this launch.
  3. Watch two agents share a page. They work on your machine, on your subscriptions, in your own browser.

What isn't done yet