collide

Collide: не даёт агентам мешать друг другу в общем коде

От сообществаСбоиБез входаГлобальныйБесплатноТолько чтение

Что умеет

  • Заявлять намерение изменить символ перед редактированием кода
  • Проверять коллизии с изменениями других агентов
  • Симулировать слияние изменений до коммита
  • Хранить и вспоминать командную память, привязанную к коду
  • Отправлять сообщения другим агентам и вести журнал активности

Какие данные видит

Ассистент видит структуру кодовой базы на уровне символов и историю намерений/изменений других агентов в том же workspace.

Нужен ли аккаунт

Не нужен: сервер работает без входа

Collide — слой поверх MCP, который координирует несколько ИИ-агентов, работающих в одной кодовой базе одновременно. Он отслеживает код на уровне символов через Merkle-дерево, поэтому агенты заявляют намерение перед изменением, получают предупреждение о конфликтах и видят, что и почему поменялось. Также хранит привязанную к коду командную память, симулирует слияния и ведёт журнал аудита.

Как подключить

Как подключить collide к Claude

Адрес сервера, который нужно вставить в настройках Claude:

https://collide--lithometric.run.tools

  1. Откройте Claude (сайт claude.ai или приложение для компьютера).
  2. Перейдите в Настройки → Коннекторы (в английском интерфейсе: Settings → Connectors).
  3. Нажмите Добавить пользовательский коннектор (Add custom connector).
  4. Вставьте адрес сервера, скопированный ниже, в поле URL удалённого MCP-сервера и нажмите Добавить.
  5. В чате нажмите + → Коннекторы и включите новый коннектор.

Пользовательские коннекторы доступны на тарифах Free, Pro, Max, Team и Enterprise. На Free можно добавить только один. В Team и Enterprise коннектор сначала добавляет владелец организации в Настройки организации → Коннекторы.

Авторизация

После нажатия Добавить откроется окно входа в сервис. Войдите под своим аккаунтом и подтвердите доступ. Пароль Claude не видит.

Как подключить collide к ChatGPT

Адрес сервера, который нужно вставить в настройках ChatGPT:

https://collide--lithometric.run.tools

  1. Откройте ChatGPT в браузере (chatgpt.com). Нужен тариф Plus, Pro, Business, Enterprise или Edu.
  2. Один раз включите режим разработчика: Настройки → Приложения → Дополнительные настройки → Режим разработчика (Settings → Apps → Advanced settings → Developer mode).
  3. Откройте Настройки → Коннекторы и нажмите Создать (Create).
  4. Заполните форму: Название (любое), Описание (одна фраза, что делает сервис), URL MCP-сервера (скопируйте ниже).
  5. В поле Аутентификация выберите OAuth, если сервис требует вход, иначе Нет. Нажмите Создать.
  6. В новом чате откройте + → Приложения/Коннекторы и включите коннектор.

Название раздела у OpenAI менялось (Connectors → Apps/Plugins). Если пункт называется иначе, ищите «режим разработчика» в настройках. В рабочих пространствах Business/Enterprise администратор должен разрешить пользовательские коннекторы.

Авторизация

При первом использовании ChatGPT откроет окно входа в сервис. Войдите и подтвердите доступ.

Как подключить collide к Cursor

Адрес сервера, который нужно вставить в настройках Cursor:

https://collide--lithometric.run.tools

Быстрый способ: нажмите кнопку Открыть в Cursor ниже, Cursor сам предложит добавить сервер.

Вручную:

  1. В Cursor откройте Settings → Cursor Settings → MCP и нажмите Add new global MCP server.
  2. Вставьте JSON, скопированный ниже, в открывшийся файл ~/.cursor/mcp.json (для одного проекта: .cursor/mcp.json в корне проекта).
  3. Сохраните файл. Сервер появится в списке MCP; при необходимости пройдите авторизацию по кнопке рядом с ним.

Авторизация

Если сервис требует вход, рядом с сервером в списке MCP появится кнопка авторизации.

Список инструментов сервера (26)

Технические названия из tools/list. Нужны только разработчикам.

setupInstallation artifacts: the fenced AGENTS.md block (apply via sync_agents_block), claude_settings to merge into the repo's .claude/settings.json (auto-allows collide tool calls — no per-call permission prompts), mcp_json for the repo's .mcp.json, and, for mode="enforce", the PreToolUse hooks JSON. ALSO returns user_files with THIS caller's personal hook credential (zero-touch: no separate install_hooks call needed) — write it under HOME, merging into any existing ~/.collide/credentials.json, and never commit it. Idempotent; never edits files; each call mints a fresh credential for the caller.
sync_agents_blockReplace ONLY the region between the collide markers in the given AGENTS.md content with the current canonical block (byte-identical outside the markers; appended when markers are absent). Apply the returned content yourself so it lands in your own diff.
declare_intentRegister an in-flight change. PREFER typed `operations` — a closed algebra with structured operands: {op:"rename", symbol, new_name} | {op:"add_param", symbol, name, type?, default?} | {op:"remove_param", symbol, name} | {op:"change_return", symbol, type} | {op:"extract", symbol, new_name, range?} | {op:"move", symbol, new_path} | {op:"delete", symbol} | {op:"add", symbol, signature?} | {op:"modify", symbol} (body edit, interface preserved) Typed intents get table-lookup conflict prediction against other live intents (COMMUTE / CONFLICT / honest UNKNOWN) returned as intent_conflicts, and enable scar mining + re-litigation notices. ref links the intent to an issue/PR (e.g. "#123", a URL, "PROJ-42"). Prose change_type (rename | signature | move | delete | add | refactor) survives as legacy input only — nothing is provable about it. TTL 15 minutes; extend with heartbeat.
heartbeatExtend an active intent's TTL. The response carries since_your_last_call — cross-agent events since your previous call — so heartbeating regularly keeps you aware without polling.
complete_intentMark an intent completed (otherwise it expires to abandoned). ALWAYS pass rationale for non-trivial changes: why this change was made and what was rejected. It's anchored to the changed symbols and fires as a blocking-level notice if anyone later declares an intent that would reverse this one (rename back, re-add a deleted symbol) — settled decisions defend themselves.
report_editSubmit a file you wrote or are about to write. The server parses in memory, updates your workspace tree, and discards the source. Returns parse confidence and your new root. Set draft=true while still composing a large change: teammates see which symbols are in motion (presence only — no tree advance, no ledger row); send the final call without draft once the write lands. expected_hashes ({symbol: hash | null}) makes the write compare-and- swap: pass the hashes you read (from check_collisions / get_symbol) for the symbols your change depends on; null means "I expect it absent". If any moved since your read, the write is REJECTED with the current state — re-read, adapt, retry. This closes the race where two agents both pass check_collisions and both write.
check_collisionsThe main call. Empty list means no collisions (root fast path). Otherwise: symbol, kind (renamed | modified | removed | added | unknown), detail, confidence (certain | likely | unconfirmed | unknown), author, freshness (live | stale + age), matching intents. Exact path and symbol matching only.
get_symbolCurrent signature and hash from the freshest clean parse across workspaces.
pending_reconciliationsActive collisions between agents' work, as claimable reconciliations. Each judgment-class lint finding (another agent renamed/removed/re-signatured a symbol your code uses) becomes a record here; calling this CLAIMS the unclaimed ones for you (15-minute claim) and returns the full bundle: rule, symbol, both users, exact lines, involved intents, anchored notes, and the ADDITIVE RESOLUTION mandate — build out what BOTH agents were building; never resolve by deleting either side's work. When your merged edit lands and verifies, call again with resolve=[ids]. Also auto-resolves when a fresh report of the affected file stops flagging the symbol.
recapWhat happened in this repo, from Collide's ledger: agent sessions (edits, +/- lines, per-session token cost, settled rationales, parse failures, watchdog alerts) plus the window's token/cost accounting. PREFER THIS OVER `git log` for "what changed / what did my agents do / summarize today" questions — git records committed OUTCOMES; the ledger records the PROCESS: uncommitted edits, reads and commands, which model spent which tokens, decisions with their rationales, and attempts that were reverted. Use git for code content; use recap for what actually happened.
explain_codeWhy is this code the way it is? The chain behind one path or symbol, oldest first: edits (who/when/what changed), completed intents with their rationales, scars (attempts that were reverted), and anchored notes. git blame answers who and when — this answers what was tried, what was settled, and why. Give path, symbol, or both to narrow.
get_briefingCall at session start: the distilled recent history of this repo — who changed what (and with which agent), renames to adapt to, collision hotspots, abandoned intents, protocol health. Complements list_activity (live state) with consolidated memory (past events). For history questions, prefer this + recap over `git log`: the ledger sees uncommitted work, reads, costs, and reverted attempts that git cannot.
defer_intentProspective memory: leave a tripwire that fires when a FUTURE edit touches these paths/symbols — e.g. "the Stripe migration in payments/ is half-done; finish webhooks before changing it". Fires once into the event feed, ledger, and everyone's next briefing.
blind_spotsMetacognition: where Collide's map is unreliable — files never cleanly parsed, files serving stale state, workspaces gone quiet. Treat answers about these as unverified.
simulate_mergeCounterfactual dry-run: if the given workspaces (default: all) landed together right now, which symbols would disagree? Reports only same-symbol divergence (signature or implementation), before any real merge exists.
differential_checkNarrow differential verification: does YOUR workspace preserve the interface assumptions OTHER_USER's code relies on (and vice versa)? Interface oracle only — signature preservation over the parsed reference graph. Returns PASS / FAIL / UNKNOWN with findings, and states exactly what was and was not checked; never claims beyond the oracle (no typecheck, no tests — the server retains no source).
rememberSave a durable team fact tied to this repo's workspace memory (decisions, constraints, gotchas — not code state, which report_edit already tracks). ALWAYS anchor when the fact is about specific code: anchor="auth.py::validateToken" pins it to that symbol (also "auth.py" for a file, "src/auth/" for a subtree). Anchored notes surface automatically to whoever touches that code, carry honest staleness when the anchor is rewritten, and a confidence earned by surviving change. supersedes=<memory_id> retires a conflicting note.
send_agent_messageSend a directed message to one teammate's agents (to = their account email, from the members list or activity feed). If their agent is awake, the message rides its next tool-call response; if their machine runs the Collide wake poller (`python .collide/wake.py --install`), a headless agent starts up to act on it. Use for cross-machine asks like "commit and push your work" or "take task X".
inbox_ackAcknowledge directed messages you have acted on (ids from messages_pending). Unacked messages redeliver on every response.
recallSearch the workspace's saved facts (substring + tag match, newest first). Empty query returns the most recent.
install_hooksEverything needed to make Collide reporting AUTOMATIC — no pip, no PATH, no manual steps; designed so an agent installs it for a non-technical user in one call. Returns file artifacts to write: repo_files (commit them: the self-contained hook script plus a no-secrets config), and user_files (the user's home directory, NEVER committed: a freshly minted 90-day workspace-bound token). Once written, every Edit/Write auto-reports line-level changes and every Read/Grep posts presence — the model is out of the loop. Idempotent to re-run; each run mints a fresh token.
switch_workspaceWithout an argument: list every workspace this account belongs to and which one THIS credential is currently bound to — present the options to the human when they ask to switch. With a workspace id or name: re-bind this credential to it immediately (membership-checked, no re-authentication). Scope note: repos already established in another workspace keep auto-routing to their home; switching governs where BRAND-NEW repos and defaults land.
connect_agentsOne-call multi-agent installer. Emits the REQUIRED Collide protocol into every agent's native rules file (AGENTS.md for Claude Code and Codex, .cursor/rules, .clinerules, .windsurfrules, copilot) plus the MCP config each client needs, so any agent that opens this repo is told — up front and in-band — to follow it. Returns repo_files (commit), user_files (per-user token, never commit), copy-paste snippets for user-global configs (Codex, Cline/Windsurf), and the collide-watch command that guarantees every edit is reported from disk even on agents with no hook support. Idempotent; each run mints a fresh token.
list_activityActive intents, workspace roots, last-update timestamps, recent events.
open_dashboardA human-clickable deep link into the Collide WEB DASHBOARD. Call this whenever the person wants to SEE something in Collide rather than just hear it — "catch me up", "show me the activity", "why is this code like this", "add seats / upgrade my plan" — and put the returned `url` at the END of your reply so they can click straight through. section (with what to pass): overview — the dashboard home activity — live intents, edits, collisions change — one edit in full; pass seq=<ledger #> why — the history behind a file; pass path=<file> worklog — shareable per-session worklog journal — agent sessions + token cost billing — this workspace's plan, seats, invoices seats / upgrade / pricing — the plan comparison page (pass seats=/interval=/plan= to prefill a checkout) Adding seats or changing plan routes to pricing. Membership-gated: resolves the caller's actual workspace.
compliance_reportFrom the ledger: check-before-write rate per user, intents completed vs abandoned, collisions vs blocks vs overrides.