This repository has moved to lithometric/collidemcp (the plugin is in plugin/). Install with /plugin marketplace add lithometric/collidemcp. This copy is archived; existing installs keep working.
Collide MCP for Claude Code: multiple Claude Code agents, one repo, no conflicts
Collide is the coordination layer for multiplayer coding agents: it keeps Claude Code, Codex and Cursor agents, on one machine or across a team's machines, from overwriting each other on one codebase, and hands each agent what the others already learned.
This is Collide's open-source Claude Code plugin: the hooks, skill and commands. Run several Claude Code agents on one repository, alone or with a team, without them tripping over each other.
- Each agent starts in the right place. When you type a prompt, the agent is handed the exact functions it is about: their signatures, who calls them, and what changed recently. Less searching.
- Agents hear about each other's work. When a teammate's agent renames a function yours uses, or starts writing the module your task needs, your agent is told on its next step.
- Writes that would clash are stopped. An edit against a function a teammate just changed is held until your agent has their version.
- Pushing takes one step.
git push is rebased onto your teammates' new commits, tested and pushed for the agent, instead of the reject-pull-retest loop.
- Notes stay with the code. What an agent said about a change it finished is kept on the functions it touched and handed to the next agent there.
In our benchmark (12 agents at once on one repository, hard tasks), agents with Collide used 53% fewer tokens than the same agents without it. The study.
Install
In Claude Code, send these as two separate messages:
/plugin marketplace add lithometric/collide-plugin
/plugin install collide@collide
That is all. The next session sets Collide up for every repository on this machine: free, no account, and nothing leaves the machine. Run as many agents as you like in as many repositories as you like; agents in the same repository see each other's work and can message each other.
/collide:status shows what Collide did on this machine. /collide:login links the machine to a Collide account: once that account's workspace is on Team, the machine joins it by itself and brings its history along, so teammates' agents on other machines join in.
What is in this repository
hooks/hooks.json: the Claude Code hooks (session start, prompt, every read, edit and command, failed calls, stop), all run through bin/collide.bin/collide: the launcher. It fetches the pinned collide-hook binary once and runs it; it fails open.skills/collide: teaches your agent how to read what Collide tells it (briefings, Δ lines, blocked writes, landings) and to say when a teammate's work changed its plan.skills/login and skills/status: sign the machine in and see what Collide is doing (/collide:login and /collide:status in Claude Code).
The MCP server is optional. To give your agent Collide's tools as well (look up any symbol, blast radius, team memory), add it once: claude mcp add --transport http collide https://mcp.collidemcp.com.
What it sends, and what it keeps
On the free version, nothing: everything stays in ~/.collide/local on your machine. A signed-in machine tells your account only its name and its totals (tokens saved, agents, repositories) until its workspace is on Team. On Team, this is what reaches Collide:
- Your code's structure: function and class names, signatures, docstrings, and which function calls which. Parsed on your machine; file contents are not sent.
- What your agents are doing: which files they read, edit and run, and when.
- Your prompts are used once to find the code they are about, then dropped. Only a numeric summary of the prompt's meaning is kept, for two hours.
Your sign-in credential stays in ~/.collide/credentials.json on your machine and is never written into the repository.
How it runs
Every hook goes through bin/collide, which runs Collide's native collide program. On first use it downloads Collide (one program: the hooks and the local server) pinned in hooks.lock for your platform, and refuses any file whose SHA-256 does not match; later updates must carry Collide's signature. Nothing runs through Python or Node. If a binary cannot run, the hooks stay silent and never block your agent.
The local server listens on 127.0.0.1 only and keeps its data in ~/.collide/local. A repository already set up with Collide's setup command keeps its own hooks and connection; the plugin stands down there so nothing is reported twice.
Plans
Free is everything on one machine: unlimited agents and repositories, no account. Team connects agents across machines, adds the dashboard and matches prompts to code by meaning (pricing).
License
This plugin (the hooks, the launcher and the commands in this repository) is MIT licensed. The collide-hook binary it downloads and the Collide service are not open source; they are free to use under Collide's terms.
Frequently asked questions
How do I run multiple Claude Code agents on the same repository?
Install this plugin (two messages, above) and start your agents as usual. They see each other's changes as they work, clashing writes are held, and pushes land in one step.
Do I need the MCP server too?
No. The hooks do the work with no extra model calls. The MCP server (claude mcp add --transport http collide https://mcp.collidemcp.com) adds tools an agent can call on purpose.
Does it work with Codex and Cursor?
This plugin is for Claude Code. For Codex and Cursor too, install Collide for the whole machine: curl -fsSL https://collidemcp.com/install.sh | sh or npx collidemcp (collide-free).
If Collide saved your agents a merge conflict, a star helps other developers running parallel Claude Code agents find it. Website: collidemcp.com · Free version and source: lithometric/collide-free
