Project Relay MCP Server
Coordinates multiple AI systems and human reviewers using Git as an inspectable record, providing tools for task management, evidence tracking, and independent review.
README
Project Relay
Project Relay is an evidence-governed coordination layer for multiple AI systems and human reviewers. It uses a Git repository as a shared, inspectable record while MCP provides a provider-neutral integration surface.
Status: pre-alpha M1 candidate. The name is a working title, the protocol is unstable, and no production deployment is implied.
Why it exists
Multi-model work usually fragments across chats, vendors, local tools, and copied prompts. Relay gives that work a common protocol:
- one explicit question per task;
- named owners, reviewers, and decision authorities;
- immutable evidence references and reproducible commands;
- independent review before acceptance;
- recorded disagreement instead of silent consensus;
- human approval for consequential decisions;
- no requirement for every model to share one vendor API.
Relay coordinates work. It does not decide whether a scientific claim is true, replace peer review, or make model output trustworthy by default.
Architecture at a glance
AI clients / humans / IDE agents
|
MCP interface
|
Relay protocol + policy
|
append-only events and evidence
|
Git repository
|
read-only web console / GitHub UI
The canonical record is validated JSON plus referenced artifacts in Git history. Issues, labels, dashboards, and the Pages console are projections of that record, never alternate sources of truth.
Current public milestone
The local M1 vertical slice now includes:
- JSON Schemas for tasks, events, evidence, reviews, and decisions;
- deterministic canonicalisation and SHA-256 provenance helpers;
- append-only event-chain verification plus policy-checked task transitions;
- an evidence-bundle CLI that hashes declared artifacts without executing commands;
- a read-only MCP server with tools, a task resource, and an independent-review prompt;
- a deterministic static console projection of state, history, and gate results;
- a synthetic submission, review, remediation, resubmission, and human-decision fixture;
- Windows and Linux CI conformance coverage plus a public-boundary scanner.
GitHub write automation, remote MCP transport, authentication, and hosted multi-tenant operation are deliberately deferred until their threat models are reviewed.
Quick start
Requires Node.js 24 or newer.
npm install
npm run check
npm run mcp
Check direct dependency versions without changing the repository:
npm run versions:check
Run the offline local doctor for measured runtime, Git, lockfile, workspace, and worktree checks:
npm run doctor
To review conservative updates and receive an interactive [y/N] installation prompt:
npm run versions:update
The MCP process uses stdio and reads the workspace specified by RELAY_WORKSPACE (the current directory by default):
{
"mcpServers": {
"project-relay": {
"command": "npm",
"args": ["--prefix", "/absolute/path/to/Project-Relay", "run", "mcp"],
"env": {
"RELAY_WORKSPACE": "/absolute/path/to/a/relay-workspace"
}
}
}
}
Use examples/m1 to inspect the complete local remediation loop. examples/minimal remains the smallest M0 kernel fixture.
Create a validated evidence bundle from an explicit manifest:
npm run evidence:create -- --manifest path/to/manifest.json --output path/to/bundle.json
The command reads and hashes declared artifacts. It does not execute the commands recorded in the manifest.
Safety and publication boundary
This public repository must never contain credentials, private datasets, unpublished ITSM material, personal information, confidential prompts, or commercial records. .gitignore is not a security boundary: releases must be exported from a separate private workspace using an explicit allowlist and reviewed as if every commit were permanent.
See Publication Boundary and Security before contributing.
Documentation
- Architecture
- Protocol and decision gates
- Public/private publication boundary
- Version maintenance
- Roadmap
- Contributing
- Licence status
Licence status
No open-source licence has been selected yet. Publication on GitHub does not grant permission to reuse the work beyond rights provided by GitHub's terms. See LICENSE-STATUS.md.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。