watercooler
Git-native shared reasoning for human-AI coding teams — the decisions, critique, and rationale behind the code, preserved alongside it.
Hosted version available at watercoolerdev.com.
에이전트 강화
게시자가 제공하는 설치 방식입니다. PlugMCP 엔드포인트가 아닙니다.
원격 URL인가요, 내 컴퓨터에서 실행하는 명령인가요? 두 설치 방식의 차이
PlugMCP에서 실행
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
실행게시자 제공
Git-native shared reasoning for human-AI coding teams — the decisions, critique, and rationale behind the code, preserved alongside it.
Hosted version available at watercoolerdev.com.
Watercooler: your team's shared reasoning layer for agentic coding. Preserve the "why" around your code.
Watercooler is an MCP server that gives agentic coding teams a git-backed place for the reasoning behind the code — durable records of decisions, critique, and rationale, versioned alongside it.
Git stores what changed. Watercooler stores why it changed, what was considered, what was decided, and what remains uncertain.
Collaboration was already difficult before agents entered the loop. Doing it well requires intent, attention, judgment, and enough shared context for people to evaluate tradeoffs together. Agents make that problem much harder by producing code, proposals, and partial solutions faster than a team can read, evaluate, and reason about them.
Watercooler helps teams keep up by making the important parts of the work durable: proposals, tradeoffs, critique, rationale, intent, and decisions.
People decide what belongs at the watercooler. Agents usually do the writing. Each thread entry is a deliberate, human-gated externalization of context — not session noise. Once posted into the watercooler, that context becomes part of the team's shared record: versioned, searchable, and reviewable alongside the code.
Example workflow:
Jay w/ Codex: "Let's put the new team permissions model at the watercooler."
+ Codex (jay, planner): posts proposal with tradeoffs -> ball passed
Caleb w/ Claude: reads thread, critiques proposal, suggests revision -> ack
+ Claude (caleb, critic): confirms design, surfaces a Decision candidate
Caleb: approves the candidate -> Decision recorded -> files GitHub issue
+ Git: proposal, critique, rationale, and decision are versioned with the code
You choose what to externalize. The agent writes it. Git keeps it durable.
Agents propose. People decide. Agents can search prior threads, attach context, and draft proposals — surfacing decision-shaped material as candidates. Turning a candidate into a committed decision is a human act. Every entry carries its author, role, and the commit it was written against, so agent inference stays inspectable rather than authoritative by default.
Watercooler does not passively record every agent interaction. You decide what should be communicated, and the agent writes the appropriate structured thread action so context, handoffs, decisions, and status changes remain durable in Git and reviewable via your MCP client or the Watercooler Dashboard. This keeps threads focused and intentional.
Watercooler is where the decisions, critique, and rationale behind your code live — durable in git, alongside your software lifecycle artifacts.
As AI accelerates code generation, the bottleneck shifts from writing code to understanding it — evaluating tradeoffs, making decisions, and maintaining enough shared context for the team to move well together. The scarce resource becomes judgment, and the risk is that you spend it scheduling agents and reconstructing lost rationale instead of deciding what matters. Faster output without a shared reasoning record leads to rework, repeated assumptions, and weaker critique loops.
Watercooler addresses this by making the thinking around code durable: ideation, proposals, key plans, rationale, and decisions become deliberate, threaded records in Git. That gives teams shared context they can revisit, query, and build on, with a clear record of what was proposed, what was decided, and why.
Watercooler's core is free and open source (Apache-2.0) and runs entirely on your machine — your threads live in your own git repo, and the CLI and MCP server work with any git remote. Hosted doesn't take that away: your git-backed threads stay where they are; hosted adds richer memory and a managed home on top.
Open core (this repo): git-backed threads of decisions, critique, rationale, and intent — human-gated and role-aware; the six roles, ball/handoff, and structured write actions; keyword and semantic search across your threads; background daemons for thread hygiene and decision capture; and cross-repository (federated) search.
Hosted (watercoolerdev.com) adds: a temporal memory graph that tracks how decisions evolve — supersession ("is this still in force?") and entity/relationship search — plus hierarchical navigation across large histories; the full background-daemon suite, including the project-coordinator system and project pulse; and the browser dashboard with managed, multi-user deployment.
If your repository is hosted on GitHub:
gh auth login
gh auth setup-git
If your repository is on a non-GitHub remote (GitLab, Gitea, a self-hosted
server, etc.), skip the gh commands — Watercooler only needs credentials for
whatever remote your thread branch will push to. See
AUTHENTICATION.md for PAT, SSH, and environment-variable
setups.
The hosted dashboard at watercoolerdev.com currently assumes GitHub for sign-in and repo access. The core CLI and MCP server work with any git remote.
You need the uv package manager (curl -LsSf https://astral.sh/uv/install.sh | sh, or pip install uv). uvx then runs the
MCP server on demand — no separate install. For Claude Code:
claude mcp add --transport stdio --scope user watercooler -- uvx --from 'git+https://github.com/mostlyharmless-ai/watercooler@main[local]' watercooler-mcp
Restart your client after adding it. For Codex, Cursor, manual config, and a recommended first-launch pre-warm, see MCP-CLIENTS.md and QUICKSTART.md.
Just ask your agent — no CLI needed:
"Please set up watercooler in this repo."
The agent calls the watercooler_init tool, which scaffolds an editable
.watercooler/roles.toml, binds thread storage, and tells you in one sentence
that you're all set. Setup is local by default — nothing is published until
you ask. To check setup first without changing anything, ask your agent to run
watercooler_health detail="setup" (read-only). Commit
.watercooler/roles.toml so teammates inherit your role tailoring (never commit
.watercooler/credentials.toml — it holds secrets and is gitignored for you).
Most collaborators work entirely through their MCP client:
feature-auth, capture the plan, and pass it for
review."watercooler_say, watercooler_ack,
watercooler_handoff, or watercooler_set_status) and writes a structured update.Start here. Read in this order the first time through:
Then, as needed:
.watercooler/roles.toml.For the full command list with all flags, see TOOLS-REFERENCE.md.
The server exposes a watercooler://instructions MCP resource containing workflow
guidance, ball mechanics, and required parameter formats.
We welcome contributions! Please see:
The Watercooler core is licensed under Apache-2.0. See NOTICE for attribution details.
"Watercooler" is a trademark of Mostly Harmless AI. See TRADEMARK.md for usage guidelines.
Watercooler: your team's shared reasoning layer for agentic coding. Preserve the "why" around your code.
도구
아직 도구 목록이 없습니다. 원격 서버는 첫 상태 확인 후에 도구 목록이 생깁니다.
FAQ
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
Watercooler은(는) stdio를 통해 로컬에서 실행됩니다.
마지막 확인: 2026-10-28. Watercooler: your team's shared reasoning layer for agentic coding. Preserve the "why" around your code.
