Hosted mex: one reviewed memory for multiplayer AI
Multiplayer AI means a team of engineers and their coding agents working from the same understanding of the codebase. Hosted mex makes that possible without anyone running the tooling: it reads your repository through a read-only GitHub App, compiles what it finds into reviewable knowledge, and serves only what your team promotes to Claude Code and Codex. When the code changes, the affected knowledge goes back to review.
The hosted beta is open
Connect a repo and try it with your team.
How it works
- Connect a repository. A person signs in, installs the mex GitHub App with read-only access, and picks a repository.
- mex compiles it. On mex infrastructure, the default branch is compiled into architecture, component, pattern and convention units, each tied to the code it describes.
- Your team reviews. Units arrive as proposals. People promote or reject them. On the first compile you can promote the whole set at once; after that, review is unit by unit.
- Agents load what was promoted. Claude Code and Codex connect through an MCP client with a project token. They load promoted units only, and can propose changes but not approve them.
- Pushes keep it current. Each push to the default branch queues a new compile. Promoted units whose grounding changed come back as drifted and wait for review instead of being served as current.
- Loads fail closed. If the latest compile is missing, running or failed, an agent's load returns that status instead of quietly serving an older snapshot.
Five screens for people
| Screen | What you do there |
|---|---|
| Projects | Connect a repository and see compile status, the commit compiled, and any error |
| Units | Browse architecture, component, pattern and convention units with the files they cite |
| Review | Promote or reject proposed units, including changes agents suggested |
| Drift | See which units drifted and the commit that changed them |
| Org | Members and invites, project tokens you can create and revoke, project deletion |
Agents do not use the dashboard. Over MCP they load the promoted units relevant to a path or task, read promoted units, and propose new units or changes.
Code access and where knowledge lives
- The GitHub App asks for repository contents and metadata, read-only. It does not request write, administration or pull request permissions, and it only reacts to pushes on the default branch.
- To compile, mex clones the default branch and deletes the clone when the compile ends, including when it fails.
- Compiled units are stored in hosted mex, kept separate per project. Hosted mex does not commit a .mex/ folder into your repository.
- Project tokens are shown once and can be revoked. Uninstalling the GitHub App stops new clones; existing units stay until an owner deletes the project.
This describes how the beta is designed to work. It is a product description, not a security certification or a contract.
FAQ
- Can we try hosted mex now?
- Yes. The beta is open. Tell us about your team and the repository you want to connect, and we will set you up.
- How is it different from the open-source CLI?
- The CLI keeps memory as Markdown in your repo and shares it through Git; you run everything locally. Hosted mex compiles knowledge from your repo on mex infrastructure, keeps it in hosted mex, and serves promoted knowledge to agents over MCP. It does not write .mex/ into your repo.
- What access to our code does it need?
- Read-only access to repository contents and metadata through a GitHub App. No write, administration or pull request permissions. The clone used for a compile is deleted when the compile ends.
- Can an agent change the shared knowledge?
- An agent can propose a change. Only a person can promote it.
- What happens if a compile fails?
- Agent loads return the compile status as an error until a compile succeeds. An older snapshot is not served in its place.
- Which agents are supported?
- Claude Code and Codex, through an MCP client configured with a project token.
Join the hosted beta
The beta is open. Tell us about your team and the repo you want to connect, and we'll get you set up.