用于有界多代理委员会会议的MCP服务器——Codex + Claude代码轮次审议
概览
Вече — 作为MCP服务器的有界多代理审议。 veche 是一个模型上下文协议（MCP）服务器，允许编排代理（Claude Code，自定义代理，你的集成开发环境）在 LLM 代理之间召开，并读取结构化的文字记录。该协议是自我消灭的：每位成员分轮回复，任何人都可以用字面令牌退出，会议在所有人在同一轮通过（或点击 max_rounds）时结束。 该名称源自中世纪斯拉夫语的вече——一个以一致掌声审议的民众集会，事项解决后解散。该集会几乎直接映射到轮回+，直到沉默。 v1 发布了两个对称的 CLI 支持适配器： - 通过Codex exec（OpenAI）- 通过claude -p（Anthropic）——带有递归守卫，防止子Claude代码重新进入同一服务器 编排者本身可能是克劳德法典，使克洛德法↔成为另一个克劳德法+法典委员会。
概述（AI 翻译）
Вече — 作为MCP服务器的有界多智能体审议。
veche 是一个模型上下文协议（MCP）服务器，允许编排代理（Claude
代码、自定义代理、你的IDE）召集委员会会议，由LLM代理和阅读
返回结构化的转录。协议是自我熄灭的：每个成员都回复
在回合中，任何人都可以使用字面令牌 退出，会议结束
即是所有人在同一轮中通过（或击中max_rounds）。
该名称源自中世纪斯拉夫语 вече——即一个由
当事情尘埃落定后，大家都欢呼并散去了。这几乎直接对应到了
rounds + until silence。
v1 发布了两个对称的 CLI 支持适配器：
- Codex 来源：
codex exec（OpenAI） - Claude Code 通过
claude -p（拟人版）——带有递归守卫，阻止了孩子 Claude 代码，因为我重新进入了同一服务器
编排器本身可能是 Claude Code，使得
Claude Code ↔ another Claude Code + Codex 委员会。架构与行为是
在 spec/ 中完全规范，使用受 C4 启发的四级模型;代码遵循
规格和更改先对规格进行。
为什么
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
Veche
Вече — bounded multi-agent deliberation as an MCP server.
veche is a Model Context Protocol (MCP) server that lets an orchestrator agent (Claude
Code, a custom agent, your IDE) convene committee meetings between LLM agents and read
back a structured transcript. The protocol is self-extinguishing: every member replies in
rounds, anyone can opt out with the literal token ``, and the meeting terminates the
moment everyone passes in the same round (or hits max_rounds).
The name comes from the medieval Slavic вече — a popular assembly that deliberated by
acclamation and dispersed when the matter was settled. That maps almost directly onto
rounds + until silence.
v1 ships two symmetric CLI-backed adapters:
- Codex via
codex exec(OpenAI) - Claude Code via
claude -p(Anthropic) — with a recursion guard that prevents a child Claude Code from re-entering this same server
The orchestrator may itself be Claude Code, enabling
Claude Code ↔ another Claude Code + Codex committees. Architecture and behaviour are
fully specified in spec/ using a C4-inspired four-level model; code follows the
spec and changes are made to the spec first.
Why
LLMs collapse into a single line of reasoning on hard problems. A standing MCP tool for pulling a second opinion from a different model — or a differently-prompted instance of the same model — broadens the search, surfaces blind spots, and produces more balanced decisions. The committee protocol drives the discussion to a stable point without the orchestrator hand-rolling turn-taking.
Install
npm install -g veche
veche install # writes SKILL.md to ~/.claude/skills/veche/ and ~/.codex/skills/veche/
# plus optional skill UI metadata under agents/openai.yaml
# plus claude mcp add / codex mcp add for the stdio MCP server
veche install --for=hermes # opt-in install into Hermes Agent (~/.hermes/skills/veche/
# plus `hermes mcp add veche --command node --args …`)
Restart your Claude Code / Codex / Hermes Agent session after installing. The skill is
invoked as /veche .
If you cloned the repo:
npm install
npm run build
node dist/bin/veche.js install
Prerequisites
| Required for | |
|---|---|
| Node.js ≥ 20.11 | running the server |
codex on PATH |
Codex participants |
claude on PATH |
Claude Code participants and as an MCP host |
CODEX_API_KEY or codex login |
Codex auth (either path works) |
claude login |
Claude Code auth |
Quick start
As a Claude Code skill
After veche install, open Claude Code and type:
/veche should we use argparse or click for a new Python CLI?
The skill first asks for launch options: whether to include a Human Participant, whether
安装
This server does not publish a one-line install command.
Open the repository installation guide 新手？先看 mcpskill.org 教程 →配置
{
"mcpServers": {
"veche": {
"command": "node",
"args": ["/abs/path/to/dist/bin/veche-server.js"],
"env": { "VECHE_LOG_LEVEL": "info" }
}
}
}