# AgentMash.dev

> Coordinates teams' AI coding agents by warning each agent which files teammates' agents are changing before it edits.

AgentMash is a coordination app for teams that use several AI coding agents, such as Claude Code, Codex and Cursor, on the same repository. It reports each agent's edits to a shared room and warns an agent before it edits a file a teammate's agent has changed. It ships as a desktop app for macOS and Windows, and as an `npx agentmash login` command-line setup that connects agents installed on a developer's machine.

## What It Is

AgentMash is a multi-agent coordination layer for team repositories. Hooks in each coding agent report edits, along with the exported symbols a change touched, to a small coordination server. Before an edit, the agent checks that server. If a teammate's agent changed the same file, or a symbol the file depends on, a graded advisory is injected into the agent's context so it can adapt mid-task. Nothing is locked: in Claude Code, Codex and GitHub Copilot the advisory is a note the agent reads and the edit proceeds. Other agents get it as a refusal of the first edit, and a retry goes through. An optional strict mode cancels edits into a teammate's change until the agent acknowledges it with a reason.

## How the workflow runs

- Agents can claim files before starting. Claims are advisory and expire.
- An MCP server lets an agent ask who is working on what while it is still planning.
- A git post-commit client reports commits from editors and agents without hooks.
- The desktop app, called lobby, runs several coding agents side by side on the user's own logins, shows teammates' sessions and activity, and lets a user send a prompt to a teammate's agent. Nothing runs until the teammate approves it, and it runs on their machine and account.
- Conflict detection matches by file path for every language. Symbol-level matching covers the JavaScript and TypeScript family only.

## The crew

AgentMash also offers 22 ready-made Claude Code agents, each with one job, such as planning, exploring, testing, reviewing, debugging, documentation, frontend, backend and database work. One command adds a chosen set to a repository as a named team, with presets for feature, bugfix, release, mobile and AI feature work. A team is files in the repository and does not need the coordination server.

## Privacy and reliability

By default the room receives file paths, branch names, line counts, the names a change touches and the first 500 characters of a prompt, not code, unless the room turns diffs on. Every hook caps its network call with a hard timeout and exits silently on failure, so a down server cannot block coding or commits. The pre-edit check has a default budget of 1500 ms. Self-hosting through a private Docker image is available on the enterprise plan. The homepage presents benchmark figures, for example fewer overlapping edits and shorter time to accepted tasks, which are vendor-published claims.

## Features
- Pre-edit advisories injected into the agent's context
- File and symbol-level conflict detection
- Advisory file claims
- Optional strict mode that blocks edits until acknowledged
- MCP server for asking who is working on what
- Git post-commit client
- Desktop app running multiple coding agents on own logins
- Live team view of online developers and agent activity
- Ask a teammate's agent with an approval step
- 22 ready-made Claude Code crew agents with team presets
- Self-hosting via private Docker image on enterprise plan
- Public API and wire protocol for custom clients

## Integrations
Claude Code, Codex, Cursor, GitHub Copilot, Devin, Amp, Antigravity, Cline, Crush, Factory Droid, Gemini CLI, goose, Junie, Kilo Code, Kiro, OpenCode, OpenHands, Qwen Code, Continue, Zed, GitHub, MCP

## Platforms
WINDOWS, MACOS, WEB, API, VSC_EXTENSION, JETBRAINS_PLUGIN, CLI

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://agentmash.dev
- Documentation: https://agentmash.dev/docs
- EveryDev.ai: https://www.everydev.ai/tools/agentmash-dev