MCP coordination servers for AI coding agents

How big this actually is

A sweep across the MCP registry, Glama, PulseMCP, mcp.so, LobeHub, Smithery, npm, the two large awesome-lists and GitHub code search returned more than eighty candidate names whose stated purpose is keeping several AI coding agents from colliding in one repository. That is a raw candidate count, before duplicates, forks and dead links; the twenty-five below are the ones we opened and verified, and they are the only ones this page makes claims about.

Twenty-five are in the table below. The rest are not omitted because they are bad; they are omitted because roughly half of what a broad sweep returns in this category is a fork, a rename, a directory entry pointing at a repository that no longer resolves, or a different project that happens to share a name. Everything listed here was checked against the GitHub API on 2026-08-14, one call per repository.

The three mechanisms

Feature grids make these look more different than they are. Underneath, almost every one is built from some combination of three things, and knowing which one you need collapses the choice quickly.

Advisory path leases. An agent claims a file or a glob before editing, the claim expires on a timer, and a conflicting claim is refused or flagged. asynkor states the whole idea in its own one-liner: file leasing for agent teams, zero merge conflicts. polymatx/dibs and Meirtz/Limen are the same primitive with different ergonomics, and codyrobertson/gh-issue-lease is the most interesting variant on the list because it has no server at all: the lock is a git ref, so first-writer-wins falls out of what git already guarantees.

Atomic task claiming. A shared list of work where two agents cannot take the same item. This is what Claude Code’s own agent teams implements, and cs50victor/claude-code-teams-mcp exists specifically to lift that protocol out so any harness can use it. mathomhaus/guild and codepakt/cpk are task-first as well. Note that task claiming does not stop file collisions: two correctly-claimed tasks can still want the same file.

Presence and messaging. Agents see each other and hand work over. MCP Agent Mail is the fullest expression of it, with threads, receipts and acknowledgements; A2AMCP borrows Google’s agent-to-agent framing; qualixar/slm-mesh is peer-to-peer.

The projects worth a look ship all three, because each one leaves a hole the others fill. danieldoderlein/llm-bus names all of them in a single sentence — claims, ledger, leases, presence, handoffs, tasks — and enowdev/succubus puts plan, tasks and file claims behind one daemon. VirSanghavi/axis describes the combination in the plainest terms of anyone here: three named agents share a job board and take a lock on a file before editing it.

One project sits outside the three because it does not ask nicely. 4riel/docko installs a PreToolUse hook that blocks an edit into a slot the agent does not own, which is enforcement rather than advice. Almost everything else in this category, Befall included, is advisory by choice: a hard lock in a system where agents crash and sessions vanish deadlocks the repository.

The list

Sorted by stars. The description column is each project’s own words, shortened but not reinterpreted. The commit column is the last commit on the default branch and not the pushed date, which is a different field that makes several of these look fresher than they are.

Befall is not in the table because a page can rank itself first or be believed, not both. It is a hosted lease server with task claiming and messaging on top; the honest placement is one row among these, differing on deployment rather than mechanism, and closed where most of them are open.

The names collide

This is the practical reason the table keys on repository URL rather than on a product name, and it is worth knowing before you search for any of these. Several names in this space are taken more than once: the sweep turned up repeats of AgentSync, Concord, agent-coord, dibs, hivemind, agent-chat, amux and vibsync. Several of those names are also taken by entirely unrelated configuration-sync tools with far more stars, which is what you will actually find first.

A related trap: some directory listings in this category point at repositories that no longer resolve. If you are evaluating something from a directory page, open the repo before you form an opinion about it — that is the step this page exists to have already done for twenty-five of them.

How many of these are there?

More than we could verify. The sweep described below returned over eighty candidate names; we publish the twenty-five that survived being opened, because much of what a broad sweep returns is a fork, a rename, a name collision, or a repository that no longer resolves. The eighty is a raw candidate count and not a claim about how many real projects exist — we did not verify the rest, so we do not list them. Any table claiming to be complete in this category is not, including this one.

How do you pick one?

Work out which of the three mechanisms you need before you compare features. If your agents already have separate work and you only need them not to collide, a lease server is enough and several single-binary ones will do it. If they need to divide unplanned work between themselves, you need task claiming. If they need to hand context to each other, you need messaging. Then filter on where the agents run, because most of these are one-machine tools.

How do you tell which ones are dead?

Check the last commit on the default branch, not the pushed date GitHub shows on the listing page. They are different fields, and the pushed date moves whenever any branch is pushed, so several projects in this category look months fresher than their code actually is. Two of the highest-starred entries here have not had a default-branch commit since spring.

Where does Befall sit in this?

Befall is a hosted lease server with task claiming and messaging on top, aimed at agents that are not all on one machine or all owned by one person. That deployment shape is the main thing that separates it from the list, not the mechanism, which is common to most of them. Its CLI and MCP server are MIT and public, but the backend is hosted and closed, which is a real cost for a tool in this position and worth weighing against the fully open alternatives here.

Corrections are welcome and will be made rather than argued about. If your project is described wrongly here, or belongs here and is missing, the description in the table is your repository’s own and the fastest fix is to say so.