Collide MCP ServerVerifiedFeatured

What is Collide MCP?

Collide MCP is a coordination server for AI coding agents. It gives Claude Code, Codex, Cursor and other MCP agents working on the same repo live, symbol-level awareness of each other's in-flight changes, plus a shared memory that flags itself when the code it describes has changed.

Each agent starts from a compact briefing built from a symbol-level map of the codebase, declares what it intends to change, checks for collisions before writing and reports edits afterwards. Collide parses reported file content in memory and discards it. What persists is symbol names, kinds, signatures, hashes, line ranges, events and the notes agents save. A free local version runs entirely on one machine, and the hosted server at https://mcp.collidemcp.com adds coordination across teammates' machines.

How to use Collide MCP?

There are two ways to start.

Local install (free, no account). Install once; it covers every repo on the machine and writes nothing into them:

curl -fsSL https://collidemcp.com/install.sh | sh

Or run npx collidemcp.

Hosted server. Connect your agent to https://mcp.collidemcp.com. Sign-in is OAuth in the browser on first use, with no keys to paste. The agent's first Collide call installs the same machine hooks, which are registered in Claude Code, Codex and Cursor user settings. Repos are found from their git remote, not configured.

Claude Code

claude mcp add --scope user --transport http collide https://mcp.collidemcp.com

Claude Desktop and claude.ai

Add a custom connector in Settings, Connectors with the URL https://mcp.collidemcp.com, then approve the browser sign-in.

Cursor

{
  "mcpServers": {
    "collide": {
      "url": "https://mcp.collidemcp.com"
    }
  }
}

VS Code

{
  "servers": {
    "collide": {
      "type": "http",
      "url": "https://mcp.collidemcp.com"
    }
  }
}

Open the dashboard at collidemcp.com to watch edits, intents and collisions live. For agents with no hook support, the collide-watch watcher (pipx install collide-watch) reports edits from disk.

Key features of Collide MCP

• A six-step protocol taught to agents in the MCP handshake: absorb a briefing, announce intent, check collisions, report edits, conclude, and merge-check.
• Typed intent operations (rename, add_param, move, delete and others) with provable commute or conflict predictions against other live intents.
• Compare-and-swap writes: pass the symbol hashes you read, and a write is rejected if anything moved since.
• Memory anchored to symbols, files or directories that surfaces automatically when an agent touches that code, and is marked stale when the anchor is rewritten.
• Hooks for Claude Code, Codex, Cursor and Hermes that report mechanically and fail open on any error, plus an optional git pre-commit gate.
• An enforce mode that blocks writes landing on a certain, breaking collision, or an advisory mode with no gate.
• Self-hosting in your VPC on the Enterprise plan.

Collide MCP tools

Every tool takes a repo_id.

• Session and awareness: get_briefing at session start, list_activity for live state, blind_spots for files the map can't trust, and compliance_report (Business plan).
• Intents: declare_intent, heartbeat (intents have a 15-minute TTL), complete_intent with a rationale, and defer_intent to leave a tripwire for the next agent.
• Checking and writing: check_collisions is the main call; report_edit submits a file after writing; get_symbol returns the freshest state of one symbol.
• Memory: remember and recall.
• Merging: simulate_merge dry-runs landing workspaces together; differential_check returns PASS, FAIL or UNKNOWN on whether your changes preserve the interfaces another workspace uses.
• Setup: connect_agents, install_hooks, setup, sync_agents_block and switch_workspace.

Use cases of Collide MCP

• Running Claude Code and Codex on one repo at once: "Check collisions before you edit auth.py."
• Avoiding duplicate work: agents declare intent so a second agent sees that a symbol is already being renamed.
• Safer merges: "Simulate merging my workspace with Dana's and tell me which symbols disagree."
• Keeping team knowledge: "Remember that validateToken must stay synchronous," anchored so the next agent sees it.
• Session handoff: start each session from a briefing of recent renames, hotspots and abandoned work instead of rereading the repo.

Troubleshooting Collide MCP

• A repo doesn't appear: hooks report a repo the first time an agent works in it, using its git remote. To watch a repo earlier, connect GitHub in the dashboard and add it.
• A repo landed in the wrong workspace: use move_repo from the agent or Settings → Repo in the dashboard.
• A write was rejected: with expected_hashes, a moved symbol rejects the write. Re-read, adapt and retry.
• Usage warnings appear: limits are soft caps. Over them, responses carry a usage_notice and agents are not blocked.
• The pre-commit gate blocks a false positive: override with git commit --no-verify.