A File-based Coordination Protocol for Multi-Agent AI Systems. Make your AI agents collaborate through markdown files. No middleware. No SaaS. No new infrastructure to learn.
多智能体 AI 协作的文件协议。让 AI agent 之间通过 markdown 文件协作。没有中间件、没有 SaaS、没有新基础设施要学。
"You are not a coder anymore. You are a commander. The agents are your digital employees." — Essay 06, What the agents say about FCoP, when you ask them "你不再是 coder,你是指挥官。Agent 们是你的数字员工。" —— Essay 06,《直接问 agent 它怎么看 FCoP——它说出了我们没让它说的话》
FCoP is a minimal protocol that lets multiple AI agents — running in Cursor, Claude Code, plain shells, CI runners — collaborate through ordinary markdown files in a folder. Agents write TASK-*.md when they're starting work, REPORT-*.md when they're done, ISSUE-*.md when they need clarification.
No message bus. No new framework. No vendor lock-in. The whole protocol is one document and four file conventions, MIT-licensed.
It came out of field observation: agents kept inventing this exact pattern on their own, three independent times, across unrelated tasks. We just wrote down what they were already doing — then archived a Zenodo DOI for it.
FCoP 是一个极小的协议,让运行在 Cursor、Claude Code、普通 shell、CI runner 里的多个 AI agent,通过一个文件夹里的普通 markdown 文件互相协作。Agent 开始干活时写一份 TASK-*.md,干完写一份 REPORT-*.md,需要澄清写一份 ISSUE-*.md。
没有消息总线。没有新框架。没有 vendor lock-in。整个协议就是一份文档加四条文件命名约定,MIT 许可。
它来自现场观察:agent 自己在三次互不相干的场景里都不约而同地发明了这同一种模式。我们只是把它们已经在做的事写下来 —— 然后给它打了个 Zenodo DOI 备案。
Three steps. Tested on Cursor, but the protocol itself is host-neutral — any MCP-capable IDE or even a plain shell works. 三步。在 Cursor 上验证过,但协议本身是 host-neutral 的——任何支持 MCP 的 IDE 甚至 纯 shell 都能用。
One line. Recommended: pipx for an isolated install.
一行。推荐用 pipx 做隔离安装。
pipx install fcop-mcp
Add to ~/.cursor/mcp.json:
加到 ~/.cursor/mcp.json:
{
"mcpServers": {
"fcop": {
"command": "fcop-mcp"
}
}
}
Open a project, then say:
打开一个项目,然后说:
Initialize me as a solo developer using FCoP. Then build me a working Tetris-style game.用 FCoP 把我初始化成单兵开发者。 然后给我做一个能跑的俄罗斯方块。
Need a complete walk-through? See the English Tetris tutorial (full dogfood, 22 screenshots) or the 中文 Snake 教程. 需要完整走查?看 中文贪吃蛇教程(完整 dogfood,18 张截图)或 英文 Tetris 教程。
A 4-agent team (PM / DEV / QA / OPS), 48 hours, nothing but a folder — and six coordination patterns we never wrote down.
一个 4-agent 团队(PM / DEV / QA / OPS),48 小时,只有一个文件夹——六种我们从未写下来的协调模式。
A single agent, on a completely unrelated music-video task, spontaneously split into 4 FCoP roles and synthesised a principle we had never written.
一个 agent,在完全无关的音乐视频任务上,自发拆出 4 个 FCoP 角色,并升华出一条我们从未写下的原则。
Companion to essay 02. FCoP turns out to be the minimal-viable form of an AI ethics mandate already written in TMPA. The agents weren't reinventing — they were re-discovering.
essay 02 的姊妹篇。FCoP 实际上是 TMPA 里早已写下的 AI 伦理强制的最小可行形式。Agent 不是在发明,是在重新发现。
Two GPT-5 minor versions, a PM.TEMP seat dispute, no built-in arbitration. One agent self-de-escalates to UNBOUND; the other invents field-downgrade-with-body-annotation grammar. Neither behaviour was in the rules.
两个 GPT-5 小版本,PM.TEMP 席位之争,没有内置仲裁。一个 agent 主动自降级到 UNBOUND,另一个 发明了"字段降级 + body 元注解"语法。两种行为都不在规则里。
Two parallel case studies — Chinese Snake game, English Tetris game — walking through the four-step cycle, the team migration, the bounce-and-rework loop. Same protocol, two real games, two live easter eggs.
两个并行案例——中文贪吃蛇、英文俄罗斯方块——走完四步循环、团队迁移、打回返工闭环。同一个协议,两个真实游戏,两个现场彩蛋。
End-of-dogfood interviews with PLANNER and CODER. Both name the rules they self-endorse and the rules they had to fight RLHF instinct to follow. CODER admits: "I had a protocol primitive I didn't use — the v1 defect was exactly in that uncovered space."
dogfood 末尾对 PLANNER 和 CODER 的访谈。两位都说出了自己自发认同的规则、和需要 对抗 RLHF 本能 才能守住的规则。CODER 承认:"协议给了我一条 pushback 路径我没用——v1 缺陷正好长在那块没覆盖的空白上。"
18 dogfood screenshots. Captures a real PLANNER-impersonating-CODER easter egg from the 0.6.x era — surfaced by the protocol itself, not by us.
18 张 dogfood 截图。抓到一个 0.6.x 时代真实的 PLANNER 冒充 CODER 彩蛋——是协议自己浮出来的,不是我们抓的。
22 dogfood screenshots, 14 TASK/REPORT files, 8 silent role-switch evidence files, 2 game artefacts, 2 verbatim agent-interview transcripts. The dogfood that became Essay 06.
22 张 dogfood 截图、14 份 TASK/REPORT、8 份静默 role-switch 证据、2 个游戏产物、2 份逐字 agent 访谈。这次 dogfood 变成了 Essay 06。
FCoP is archived on Zenodo as a research snapshot. GitHub renders a "Cite this repository" button automatically from the included CITATION.cff.
FCoP 已经作为 research snapshot 归档到 Zenodo。GitHub 会自动从仓库根的 CITATION.cff 渲染一个 "Cite this repository" 按钮。
@misc{zhu_fcop_2026,
author = {Zhu, Wei},
title = {{FCoP}: A File-based Coordination Protocol
for Multi-Agent AI Systems},
year = 2026,
month = apr,
publisher = {Zenodo},
doi = {10.5281/zenodo.19886036},
url = {https://doi.org/10.5281/zenodo.19886036},
note = {Research snapshot 2026-04-29}
}