Collide MCP

NewTeam is free for 14 days, no card needed

Multiplayer for AI coding agents.

Every agent on your team sees what the others are doing, so they stop colliding and stop re-reading.

47% fewer tokens with 12 agents, measured ·set up in a minute: one URL or the Claude Code plugin

reading checkout.py

maya · Claude Code

renaming total()

sam · Codex

renaming total()

lee · Cursor

fixing the export

ana · Claude Code

renaming a helper

jo · Gemini

Five agents, one repo. Sooner or later, two reach for the same code.

With Collide

Every agent on one screen, live

Who is on which file, what they are about to change, and who is heading for the same code, before anyone writes it.

Works with anything that speaks MCP

See.

hooks

Every agent's position, intent and edit, symbol by symbol, as it happens. Reported by the harness itself, so nothing goes unobserved.

Brief.

get_briefing

Before the agent asks: who is on what, what changed and why, what the tests said. Exact signatures, not summaries, in a couple of thousand tokens.

Steer.

plan_work

Work that would collide is held for the seconds it takes the other change to land, then released. The collision never gets written.

Git finds conflicts at merge time. By then the damage is already written.

One agent renames a function while another writes new code that calls the old name. Both merge cleanly, and the project is broken anyway. Four agents do an afternoon of work in ten minutes, so that blind window is no longer survivable.

Connect your agents in about a minute.

Get started

The protocol

Four verbs. Zero collisions.

Most coordination tools serialize agents with file locks. Collide keeps everyone working in parallel and makes the work mutually visible. Four verbs, shipped to every agent in the MCP handshake, with zero setup.

• get_briefing
• declare_intent
• check_collisions
• report_edit

1. 01

   Absorb the week in one call

   get_briefing At session start the agent pulls the distilled recent history: renames to adapt to, collision hotspots, half-done work left behind.
2. 02

   Announce the change

   declare_intent Before starting, the agent posts what it plans to touch. Intents expire on their own if the work stalls, so stale claims never pile up.
3. 03

   Check before writing

   check_collisions One call answers: has anyone changed what I am about to depend on? Exact symbols, the new signature, who did it, how fresh.
4. 04

   Report the edit — or let the hook

   report_edit The server parses the file in memory, diffs the changed lines, updates that agent's tree, and discards the source. With hooks installed, this happens on every single write, automatically.

Token costs

47% fewer tokens with twelve agents on one repository

Most of a coding agent's bill is re-reading: every search, every read and every retried push is another message that replays the whole conversation. Collide removes those steps. Same model, same prompts, same day: 7.8M tokens with Collide, 14.7M without.

5.4 vs 10.4

searches before the first edit

It starts in the right file

Your prompt is matched to the code by meaning, so the agent gets the right functions, their signatures and callers before its first search.

14 vs 25

turns per agent

It hears about teammates as it happens

When a teammate's agent changes something yours uses, your agent is told on its next step. No re-reading to find out what moved.

12/12

tasks correct with 12 agents at once

Writes that would collide wait

An edit against something a teammate just changed waits for their version. Nothing is written twice.

4.0 vs 8.9

git calls per agent

Pushing is one step

git push is rebased, tested and pushed by the hook, not the model. The retry loop used to be half an agent's calls.

Read the study · All benchmarks

Benchmarks

Measured on real runs, not estimated

Same task, same model, same 300-module repo. The only difference is whether the agent had to discover the codebase itself, or was handed exact facts before it asked. Every agent in every round landed the change correctly. Two ways to count it below: the arithmetic the field quotes, and the bill.

By Graphify's arithmetic

621×

1,182,646 tokens an agent spent discovering the repository on its own, over the 1,904 tokens of facts a briefing hands it instead. Raw over compact, per query, the way the 71.5× is computed.

vs

What you are billed

5.3×

Study 2: the whole session, per agent, ten agents each way, same task, same model, same run, every one correct. The agent still edits, runs the tests and reads what it must; the facts are the small part. This is the number we publish.

Tokens per agent · same task

No Collide — the agent greps for itself 1,183K 26.8 messages · 6 agents Hooks brief the agent before it asks 420K 9.2 messages · 6 agents …plus the repo's shape up front 356K 7.8 messages · 5 agents …plus what changed, and when the list is complete 301K 6.7 messages · 7 agents …plus replaying work another agent verified 194K 4.8 messages · 10 agents

5.3×

fewer tokens per agent than agents searching the repo themselves, same run

100%

tasks landed correctly, every round

21/21

features correct against the live server

The method, the raw numbers, and what they don't say →

The radar

Precision beats recall

Each person's bubble says what their agent is doing: the change it claimed, the collision it was held off, what comes next. Every finding names who, which agent, the exact symbol and lines.

When Collide speaks up, it means it: a half-typed rename is never reported as a deletion, and red is kept for what actually happened.

Connect

One URL, any client

Paste it as a connector in claude.ai or ChatGPT, or add it once from the CLI. Every credential is bound to one invite-only workspace.

https://mcp.collidemcp.com

One-click

Claude ChatGPT Cursor VS Code

1. claude.ai · Settings → Connectors → Add custom connector → paste the URL (auth auto-detects; no headers, no client ID)
2. ChatGPT · Settings → Apps & Connectors (developer mode) → Create → paste the URL → auth: OAuth
3. Cursor · VS Code · the buttons above, or add a remote MCP server with this URL
4. Codex · Windsurf · Zed · Gemini CLI · add a remote (HTTP) MCP server pointing at this URL
5. Anything else that speaks MCP · same URL, remote server — no per-client setup

Approval opens in the browser; sign in, pick a workspace, done. No keys, no headers, no client IDs.

Then, in your agent, say setup collide — it wires this repo up (AGENTS block + auto-report hooks). It also runs on its own the first time, so this is just in case.

Open source

Collide's hooks and skill for Claude Code, MIT licensed

The plugin that connects Claude Code to Collide is open: the hooks that report each read, edit and command, the launcher that runs them, the skill that teaches an agent to read what Collide tells it, and the commands. No MCP setup needed; sign in once per repository from the terminal.

GitHub: lithometric/collide-plugin What is in it

# in Claude Code, one message each

/plugin marketplace add lithometric/collide-plugin

/plugin install collide@collide

# in the repository to connect (a browser sign-in)

/collide:login

For teams

The moment there's more than one of you

The moment a second person or a second agent touches the repo, Team puts every agent and every person on the same radar.

Everyone on one radar

Every teammate's agents report into a single invite-only workspace. You see who — and which agent — is touching what, live, across the whole team.

Parallel, never serialized

No file locks, no waiting your turn. Collide makes concurrent work safe instead of blocking it — the whole point is that four agents can run at once.

Onboard in one paste

Invite by link, bound to one workspace. A new teammate pastes the URL, picks the workspace, and is productive — no keys, no setup, no config review.

Start Team

Team $19/seat/mo, free for 14 days, no card · invite a teammate and you both get 14 days of Team · teammates can buy their own seat

Viewers are free and unlimited — they watch the radar and read the briefings. Players, whose own agents report in, are the seats.

Compare the plans

Plans

When the ledger has to be provable

Team covers most teams. Past it: compliance, then self-hosting and audit.

Business

compliance_report

$45/ seat · mo

When the ledger stops being a feed and becomes an artifact.

Everything in Team, plus the governance a growing org needs: the full history kept forever, sign-in you control, and a per-person compliance view that catches a regressed check-before-write habit before the collisions do.

• Unlimited history · unmetered calls
• Compliance: the full ledger of every agent action, audit logs, export
• compliance_report — check-before-write rates per person & agent
• SSO, admin roles and permissions
• Priority support See Business

Enterprise

self-hosted

Customtalk to sales

Your VPC, your data-residency rules, the ledger as an audit trail.

Everything in Business, run the way your security team needs it: the exact same server inside your own VPC or on-prem if you want it, the append-only, hash-chained ledger as an exportable audit artifact — and an unattended reconciliation runner that adapts your agents' work to teammates' changes for you.

• SAML or SCIM provisioning
• Custom data retention
• Security review packet with a DPA
• Private or dedicated deployment (VPC) · autonomous reconciliation runner
• Uptime SLA and a dedicated contact Talk to sales

Every price is grandfathered for life, and every limit is a soft cap — Collide warns and keeps serving, never blocks. Full pricing →

FAQ

Asked and answered

▸How does Collide cut AI coding agents' token costs?

By removing the steps agents spend re-reading. Each prompt is matched to the code it is about, by meaning (code embeddings, ranked by the code graph so used code beats lookalikes), and the agent is handed those functions' signatures and callers before it searches. Teammates' changes arrive as they happen instead of being rediscovered, and git push is rebased, tested and pushed by the hook in one step. In a measured run of twelve concurrent agents on one repository, agents on Collide used 47% fewer tokens, every task correct.

▸What is the Collide MCP server URL?

https://mcp.collidemcp.com. In Claude Code: claude mcp add --transport http collide https://mcp.collidemcp.com. In claude.ai or ChatGPT, add it as a custom connector. Or skip MCP entirely and install the open-source Claude Code plugin, which connects through hooks.

▸Is Collide open source?

The Claude Code plugin is: its hooks, launcher, skill and commands are MIT licensed at github.com/lithometric/collide-plugin. The native hook binary it downloads and the Collide service are free to use but not open source.

▸What is Collide?

Collide is an MCP server that gives AI coding agents — Claude Code, Codex, Cursor, and anything else that speaks the Model Context Protocol — live, symbol-level awareness of each other's in-flight changes. Agents check before they write, declare what they're about to change, and get told the moment a teammate's edit collides with theirs.

▸Which coding agents does it work with?

Any MCP client gets the full protocol automatically. Claude Code and Codex additionally get automatic edit reporting and a pre-write blocking gate via native hooks; agents with no hook support are covered by a filesystem watcher, and a git pre-commit gate protects every commit from any agent or human.

▸Does my source code leave my machine?

Reported file content is parsed in memory and discarded. What persists is structural metadata only: symbol names, signatures, hashes, line ranges, and events. Live code snippets in the dashboard feed expire from a time-limited ring and are never written to the durable ledger.

▸What happens when two agents edit the same function?

Collide detects the divergence at the symbol level and tells both sides who changed what, with a confidence rating. Writes can be made compare-and-swap: if the symbol moved between an agent's read and its write, the write is rejected with the current state so the agent adapts instead of clobbering.

▸How much does Collide cost?

Viewers are free and unlimited: invite anyone and they watch the radar, the activity feed and the briefings. Players — people whose own agents report in and get briefings and collision checks — are the seats you pay for. Free is one player (you), one repo, up to ten agents running at once, messaging between your own agents and a 7-day history, on the same engine as every paid tier; invite a teammate and you both get 14 days of Team. Team is $19 per seat per month ($190 per seat per year) from two seats: unlimited agents and repos, agent knowledge shared across teammates, messaging across the whole team, 90 days of history and the team dashboard, with a 14-day free trial and no card required. Business is $45 per seat per month ($450 per seat per year) with unlimited history, compliance export and audit logs, SSO and admin roles. Enterprise is custom: SAML or SCIM, custom retention, a DPA, a private or dedicated deployment and an SLA. Every price is grandfathered forever.

▸Can a teammate pay for their own seat?

Yes, the way players buy their own copy of a multiplayer game. Once the owner's workspace is on Team, a viewer can buy their own seat at the same $19 per month ($190 per year) and becomes a player; their seat uses none of the owner's. The owner decides in settings: approve each request (the default), let anyone buy in, or turn it off.

▸Can I self-host Collide?

Yes — Enterprise deployments run the same server in your own VPC or on-prem, with custom retention, data residency, audit logging, and the hash-chained ledger as an exportable audit artifact.

Still deciding? Setup takes about a minute and the free tier has no time limit — you can point it at a real repo this afternoon.

Get started

col · lide

Put your agents on the same team.