Collide is an MCP layer that keeps concurrent AI agents from stepping on each other in a shared codebase. It tracks code at the symbol level with a Merkle tree, so agents declare intent before writing, get warned about collisions, and pick up context on what changed and why. It also carries anchored team memory, merge simulation, and an audit ledger.
Scored across 14 tools
Most tools target clearly distinct operations, and descriptions proactively cross-reference each other (e.g., blast_radius vs get_symbol, briefing vs explain_code), which strongly aids selection. The main overlap is setup vs connect_agents, where connect_agents explicitly returns 'the same one install command as setup', and the cluster of six read-only info tools (get_briefing, explain_code, recall, recap, get_symbol, blast_radius) requires careful reading to pick correctly.
All names are lowercase snake_case, which is consistent casing, but verb conventions are mixed: verb_noun (get_symbol, explain_code, defer_intent), bare verbs (claim, recall, remember, setup, recapit), and noun phrases (blast_radius, pending_reconciliations, inbox_ack). The mix is readable but does not follow a single predictable pattern.
14 tools sit at the upper end of a reasonable range for a server spanning code graph queries, team notes, messaging, coordination primitives, and setup. Each tool has a discernible role, though setup/connect_agents borders on duplicative and could likely be consolidated, keeping it just shy of a clean 5.
Coverage across the apparent domain is solid: read paths (get_briefing, get_symbol, blast_radius, explain_code, recap), notes lifecycle (remember + recall), messaging (send_agent_message + inbox_ack), coordination (claim, defer_intent, pending_reconciliations), and setup are all present. Minor gaps exist—no explicit tool to edit/delete a note or list pending messages independently of the hook-driven message stream—but staleness marking and hook-based delivery make these workable.
Before renaming, changing a signature or deleting: every symbol that depends on this one, level by level, with each dependent's file, owner and any in-flight work over it. Answers "what breaks if I change this" from the live code graph; skip it when the briefing's Coverage line says every dependent is already listed. Read-only.
Inputs: path (required), depth, symbol, repo_id.
Reserve the next number in a sequence (a migration, an ADR, a release) so no other agent takes it; two agents asking at once never get the same number. The hooks already claim a numbered file's number when you create it, so call this only for a number you need before writing, or for a named sequence. Returns the number, zero-padded like the files already there.
Inputs: sequence (required), after, title, repo_id.
Set up Cork for the agent tools on this machine (Claude Code, Codex, Cursor, Hermes): returns the same one install command as setup, which installs Cork's hooks once for every repo on the machine and writes nothing into the repo. mode "files" returns the protocol rules files and MCP configs to commit for tools without hooks (Windsurf, Cline, GitHub Copilot) instead.
Inputs: repo_id (required), mode.
Leave work unfinished safely: a tripwire on paths or symbols that warns whoever next edits them (for example "webhooks in payments/ are half-done; finish them before changing this"). It fires once into their next tool result and briefing, and becomes a task in the team's CRM when one is connected. Returns the tripwire id and expiry.
Inputs: note (required), paths, repo_id, symbols, +1 more.
Why the code is the way it is: the history behind a path or symbol, oldest first — each edit (who, when, what changed), completed intents with their rationales, reverted attempts and anchored notes. It includes uncommitted work and the reasons given. Read-only.
Inputs: path, symbol, repo_id.
Only when no Cork briefing is in your context (Cork's hooks are not installed): the recent history of this repo — who changed what, renames to adapt to, hotspots, abandoned work and who is working right now. Where the hooks run, the same briefing already arrived with your prompt; do not call this again for it. Pass path (and symbol) to also get the blast radius of what you are about to change. Read-only. Returns summary lines, changed modules with exact signatures, anchored notes and live agents.
Inputs: days, path, depth, symbol, +1 more.
Exact facts about code without opening the file: the signature, parameters, calls, span and hash of one symbol, or, omitting symbol, of every symbol in the module, with the notes anchored to it and whether the file passed the repo's check. Use it when a briefing lacks a fact; use blast_radius for who depends on it. Read-only.
Inputs: path (required), symbol, repo_id.
Mark messages from teammates' agents as handled so they stop arriving: pass the ids from messages_pending once you have acted on them. Unacknowledged messages repeat on every response. Returns the messages still pending.
Inputs: ids (required), repo_id.
When a response says two agents' work collides (reconciliations_pending): claim the open collisions, with both sides' intents, changed lines and notes, to merge them. Merge by keeping both sides' work; never by deleting either. After the merge passes the tests, call again with resolve=[id]. Returns the claimed reconciliations and how many stay open.
Inputs: repo_id, resolve.
Search the team's saved notes (made with remember) by text and tags, newest first; an empty query returns the most recent. The hook briefing already carries the notes anchored to the code you touch: use this only to search beyond them, before re-deciding something the team may already have settled. Read-only. Returns each note with its anchor, confidence and whether it may be stale.
Inputs: tags, limit, query, repo_id.
What each agent did in this repo, session by session: edits, lines, tokens, model, estimated cost, rationales and top paths. Use it to answer "what happened" or "what did this cost". Read-only. Returns one entry per session and a dashboard link.
Inputs: days, repo_id.
Save a durable team fact: a decision, constraint or gotcha, not code state. Anchor facts about specific code (anchor="auth.py::validate_token", "auth.py" or "src/auth/"): anchored notes reach whoever touches that code and are marked stale when it is rewritten. When the human states a standing decision ("remove X", "never Y", "only Z"), save it with kind="decision" anchored to the code it covers: it never goes stale, and each agent's first write there is held once with it. Writes the note(s) and returns their ids.
Inputs: fact, kind, tags, facts, +3 more.
Send a message to teammates' agents. In a shell, cork message "..." does the same with no tool to load. With neither to nor anchor, it goes to the agents at work in this repo whose files the message names (your own other sessions included), and to everyone at work only when it names no code. to = one account email; or anchor = the code it is about ("path.py::symbol", "path.py" or "dir/") with no to, which reaches whoever has an open intent on, or edited in the last day, that code or its callers. With an anchor the recipient also sees whether the code changed since you sent it and whether its tests have passed since. It arrives with their next tool result. to = "@user" (or "@user/2", one of their agents) messages any Cork user once they accept you as a contact; hub = a hub's name or id (a new name with to="@a, @b" starts it). Returns the message id and who received it. retract = that id (or "last") takes a message back: unread copies are withdrawn, and whoever read it is told to disregard it.
Inputs: to, hub, anchor, message, +2 more.
Set up Cork on this machine, once for every repo on it. The answer has install.command: run it exactly as given, in a shell at the repo root. It installs Cork's hooks for Claude Code, Codex, Cursor and Hermes, links the machine to your account, and writes nothing into the repo; if the repo still has Cork's old per-repo setup (.collide/ hook files), it removes that and stages the removal for you to commit. Returns the command and what it installs.
Inputs: repo_id (required), mode.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Add one secure layer between your agents and this server.
