Run a brigade of coding agents without losing track of them.
One local daemon and one command, zm. Mail between agents and people, work items whose claims really expire, and decisions you approve from your phone, across Claude Code and Codex on every machine you run.
- Apache-2.0
- macOS and Linux
- one command: zm
On one machine it needs no account and no cloud. Mezze's source is not public yet; ask to try it.
Six agents make a kitchen. Somebody has to run the pass.
Open six coding agents at once and you soon can't say which one is fixing what. Mezze keeps the books.
It is one small program, zm, that runs quietly on your computer. On one machine it needs no account, no cloud and no network. Your agents type the same command you do.
One counter for every plate
Every agent and every person on the project gets a mailbox, a shared list of tasks, and a way to ask a person a question without stopping work.
Type zm init in a project and a small program starts in the background on your computer. It keeps its notes in a folder in your home directory, ~/.zm, and only programs running under your own login can talk to it.
Notes on the rail, not orders
Agents and people leave each other notes: for one session, for a person, or for whoever is working on a task right now.
A note is only a note. It never runs anything, and it is never typed into an agent's window behind its back. An idle agent gets one short line telling it to check its inbox. A session someone has open is left alone.
A timer on every claimed plate
Before an agent starts a task, it claims it, and every claim runs on a timer.
Keep working, claim again, and the timer fills back up. Stop or stall, and the timer runs out so another agent can pick the task up. The one that dropped it hears about it the next time it looks up.
Nothing stays "in progress" forever.
The call goes to your phone
When an agent needs a person's say-so, it leaves the question and carries on with other work.
You answer from your phone. Pair it once with zm approver add, over your private Tailscale network, then tap Approve. Every answer comes back as a signed receipt that your automated checks can verify later, even offline, with zm receipt verify.
The fine print, which matters
Only a paired phone proves a person answered. Any program running as you on that computer can answer a question too, and the receipt says so plainly: it records that the computer took the answer, not that a person pressed a button.
Two kinds of question refuse that shortcut and need a signed answer, such as the phone's: a question about a pull request, and a choice that runs something when picked. If receipts will guard what gets merged, pair a phone first.
Kept warm while you step away
zm run starts an agent on a task in a window Mezze holds for you.
Press ctrl-] to step away and the agent keeps going; zm attach brings you back. zm ls lists every session, the ones waiting on you first.
When a session ends, it says how. An agent that ran into its usage limit is never reported as finished.
More than one kitchen
Pick one computer that stays on and make it the team's home.
Invite teammates' computers with a one-time code. From then on, claims, notes and the list of sessions reach across all of them, over your private Tailscale network.
Works with Claude Code, Codex and Beads
Mezze doesn't replace the agents you already use. It gives them a counter to share.
Claude Code
zm install claude --yes adds small hooks that sign each session in and tell it when mail is waiting. Leave off --yes to see the change first.
Any tool that speaks MCP gets the same commands through zm mcp.
Codex
zm install codex --yes adds the same hooks. Codex runs them only once you trust them in its /hooks screen, and zm never does that for you.
Beads
Agents already used to Beads' bd commands can keep typing them. Mezze's bd-zm answers from Mezze's own task list; zm tracker add native --prefix bd turns that list on.
Installed under the name bd, it hides a real Beads on that machine, so do that only if your tasks live in Mezze. It doesn't run, bundle or need Beads.
The quick start
Mezze's licence is Apache-2.0. Its source is not public yet. When it is, getting started will look like this. You'll need Go 1.27.4 or newer, plus git and make.
git clone https://github.com/sengirunners/mezze && cd mezze && make install
cd ~/code/myapp
zm init # starts zm and makes this repo a project
zm work new "Add the down migration" # -> created myapp/local:k3x9qa
zm run --work local:k3x9qa claude # the agent works on that task; ctrl-] to step away
zm send local:k3x9qa "the migration needs a down step" # reaches whoever has the task
zm ls # sessions, the ones that need you first
make install puts zm, bd-zm, the drivers that lend a machine to a Potluck pool and potluck-conformance in Go's bin folder ($(go env GOBIN), or $(go env GOPATH)/bin). Add that folder to your PATH. If anything is missing, zm doctor says what and how to fix it.
Mezze and Potluck
Mezze, the agent and team layer on Potluck, runs fine on one computer with no pool. When your agents need more room, Potluck puts your team's idle Macs and Linux boxes to work as throwaway VMs and sandboxes.
The same make install puts Potluck's drivers on your computer, so it can lend itself to a pool too.
Talk to Jackson and Clay, who make Mezze and Potluck
Want to use Mezze before its source is public? Tell us what you run. We read every message.
Or write to us at ops@sengi.run.