Marchward MCP

Name: Marchward MCP
Author: marchward

作者：marchward版本 0.2.4i★ –i在你的电脑上运行

注册名称ai.marchward/mcp-server

安装

概览

作者提供的说明（英文原文）

Runtime authority for AI agents: credential mediation, spend cap, approval gates, audit log.

PopNx MCP 的三大支柱

这三项永远不会合并成一个分数：意见、测量和人气是不同的东西。

用户评价i暂无评价评价功能将在后续阶段与账户功能一起开放
我们的基础检查i1 项中通过 1 项依据公开信息检查；尚未扫描代码
使用情况i尚未统计★ –

它能帮你做什么

我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。

连接方式: 在你的电脑上运行
传输方式: stdio
分类: 其他
你需要填写：MARCHWARD_API_URL, MARCHWARD_API_KEY

安装

选择你使用的 AI 应用，然后复制设置。

安装之前

基础检查：1 项中通过 1 项
在你的电脑上运行，拥有你的用户账户的权限（文件、网络、应用）。
需要你自己填写： MARCHWARD_API_URL, MARCHWARD_API_KEY

标为机密的值是你账户的钥匙。千万不要分享给别人，也不要粘贴到聊天里。

你需要

npm已安装 Node.js（用于 npx 命令）

你用的是哪个应用？

ClaudeClaude CodeCursorVS CodeChatGPT

在上方选择你的应用。我们会记住它，下次按钮会显示“安装到 …”。

打开 Claude Desktop → Settings → Developer → Edit Config，把下面的内容粘贴到 claude_desktop_config.json

{
  "mcpServers": {
    "marchward": {
      "command": "npx",
      "args": [
        "-y",
        "@marchward/mcp-server"
      ],
      "env": {
        "MARCHWARD_API_URL": "your-value",
        "MARCHWARD_API_KEY": "your-value"
      }
    }
  }
}

在终端中运行这条命令

claude mcp add marchward --env MARCHWARD_API_URL=your-value --env MARCHWARD_API_KEY=your-value -- npx -y @marchward/mcp-server

运行前请先填写空着的值。

在 Cursor 中打开

或者自己复制设置

粘贴到项目中的 .cursor/mcp.json，或 ~/.cursor/mcp.json

{
  "mcpServers": {
    "marchward": {
      "command": "npx",
      "args": [
        "-y",
        "@marchward/mcp-server"
      ],
      "env": {
        "MARCHWARD_API_URL": "your-value",
        "MARCHWARD_API_KEY": "your-value"
      }
    }
  }
}

浏览器会先询问是否打开应用，然后应用会显示设置并请你确认。密钥的值会留空，由你自己填写。

在 VS Code 中打开

或者自己复制设置

粘贴到项目中的 .vscode/mcp.json

{
  "servers": {
    "marchward": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@marchward/mcp-server"
      ],
      "env": {
        "MARCHWARD_API_URL": "your-value",
        "MARCHWARD_API_KEY": "your-value"
      }
    }
  }
}

浏览器会先询问是否打开应用，然后应用会显示设置并请你确认。密钥的值会留空，由你自己填写。

这个服务器在你的电脑上运行，而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。

你的选择通过 Cookie 保存在这台设备上，无需账户。

需要你填写的值

MARCHWARD_API_URL 必填 env
Marchward API base URL
MARCHWARD_API_KEY 机密必填 env
Your Marchward API key (mw_...)

这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同，请以 README 为准。

它能做什么

它在你的电脑上运行；安装后，工具列表会显示在你的 AI 应用中。

基础检查i

我们检查的是任何人都能看到的信息；我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月6日

✓ 通过
作者与仓库一致
注册名称已通过域名 marchward.ai 验证（注册中心会核实域名所有权）
… 尚未检查
仍在维护
我们还没有从 GitHub 读取这个仓库
… 尚未检查
有许可证
我们还没有从 GitHub 读取这个仓库
– 不适用
端点有响应
它在你的电脑上运行，没有互联网端点
– 不适用
使用 OAuth 登录
在你的电脑上运行，改用本地设置
… 尚未检查
工具描述中没有隐藏指令
它在你的电脑上运行；我们只能读取互联网端点的工具列表

与“其他”中的其他服务器比较i

星标数多于此分类中其他 32505 个服务器的 0%
发布到注册中心的时间比此分类中其他 32505 个服务器的 14% 更新
基础检查结果等于或优于此分类中其他 32505 个服务器的 100%

与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时，星标数才计入。

常见问题

Marchward MCP 安全吗？

它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前，请先看看这个服务器要求哪些权限，并使用只具备所需权限的密钥。

它需要哪些密钥或代码？

你需要填写：MARCHWARD_API_URL, MARCHWARD_API_KEY

它在我的电脑上运行还是在云端？

它在你的电脑上运行。你的 AI 应用会在需要时启动它，数据不会经过作者的服务器（除非它自己调用在线服务）。

它还在维护吗？

GitHub 仓库最后更新于–，最新版本（0.2.4）于4个月前进入注册中心。

哪些应用可以使用它？

支持 MCP 的应用，例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Node.js。

有什么可以替代它？

同一分类中有：Spec Driven Development, Remote Desktop Commander, topo^alpin

替代选择

同一分类中其他作者的服务器，按星标数排序。

如何看卡片：★ GitHub 星标i三大支柱：评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i

Spec Driven Development

作者：formulahendry

MCP Server that facilitates spec-driven development workflows, not just Vibe Coding.

在你的电脑上运行★ 440MIT14个月前更新

安装

Remote Desktop Commander

作者：desktop-commander

Remote Desktop Commander: use files, terminal and processes on your own computers from any AI app.

在云端运行★ 211昨天更新

安装

topo^alpin

作者：topoalpin.app

Avalanche bulletins and mountain weather forecasts for any place in the Alps.

在云端运行今天更新

安装

作者的 README

打开 README

按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。

Marchward · runtime authority for AI agents

Marchward

Govern what your AI agents are allowed to _do_. Marchward is one open layer that sits in front of an agent's tool calls and enforces four controls before any action runs, then writes a tamper-evident record of every decision. Self-hostable, Apache-2.0, framework-agnostic.

Credential mediation. The agent holds one Marchward key, never your real GitHub, Stripe, or database credentials. The real secret is injected at execution, and the agent never sees it.
Inference spend cap. A hard ceiling on agent spend, so a runaway loop can't drain a budget.
Approval gates. Irreversible actions pause for a human instead of running unattended.
Tamper-evident audit. Every decision appends to a hash-chained log you can verify offline. Provable, not just logged.

Most tools in this space cover one of these. Marchward brings all four together in one place, open and self-hostable.

Output guardrails check what an agent _says_ (toxicity, PII, topic). Marchward governs what an agent _does_ (spend, delete, act). A check on the agent's output cannot stop an agent that still holds the keys and can make a different call. Marchward governs the action itself.

Built for teams already running agents in production (LangGraph, LangChain, raw SDKs, any MCP server, or one HTTP call from any runtime) who need an agent to be autonomous without being able to drain a budget, take an irreversible action unattended, or act without an audit trail.

Quickstart: govern an agent in 60 seconds (fully local, no account)

marchward-proxy wraps any MCP server as a child process and evaluates every tools/call against your policy before it runs (block, allow, or pause for human approval), entirely on your machine, with your own credentials.

# Write a policy (escalate anything destructive, allow the rest)

cat > policy.json <<'EOF' { "rules": [ { "match": { "toolPattern": "*delete*" }, "decision": "ESCALATE" }, { "match": { "toolPattern": "*" }, "decision": "ALLOW" } ]} EOF

Put Marchward in front of your agent's MCP server, local mode, no API

npx -p @marchward/proxy marchward-proxy \ --command "node my-mcp-server.js" \ --policy-file ./policy.json \ --local

Every tool call your agent makes now passes through policy first, and lands in a local tamper-evident (hash-chained) audit log you can verify offline. No account, no network call to us.

Prefer the SDK? pip install marchward and wrap calls with marchward.execute(). Point it at your own self-hosted setup, or at the managed plane (below).

What's open vs. what's hosted

The open packages in this repo are a fully functional self-host governor. Some operational features need shared state and run on the hosted plane (free tier, no card):

	Open (self-host, this repo)	Hosted plane (Marchward Cloud)
Execution control: allow / block / escalate tool calls by policy	yes	yes
Local governor (MCP/HTTP intercept, your own credentials)	yes	yes
Tamper-evident audit: hash-chain primitive + offline verify	yes	retained, queryable, monitored
Approval-gate decision	yes	+ managed workflow (notify, approve, resume)
Inference cost cap	local accounting	enforced (rolling-window spend)
Credential mediation (agent holds one key, real cred injected, the agent never sees it)	yes, from your own local secret store	yes, managed vault
Managed credential vault, multi-tenancy, SSO/RBAC	no	yes

So: self-host this for execution control, credential mediation from your own local store, a local cost cap, and a local audit log; add the free hosted tier for the managed credential vault (rotation, team sharing), approval workflows, retained audit-as-a-service, and cross-agent enforcement. Same engine underneath. The proxy's --remote mode just points at the plane.

How it works

Your agent makes a tool call (MCP, HTTP, or via the SDK).
Marchward evaluates it against your policy. Risk is judged by the resolved action: any destructive operation (a DELETE, a flagged destructive POST) is gated regardless of what the tool is _named_, so a custom-named destructive tool cannot slip past.
The decision, ALLOW / ESCALATE / BLOCK, is enforced before the call runs and appended to a hash-chained audit log.
Credential mediation means your agent holds only a Marchward key. The proxy injects the real service credential (from your own local store when self-hosted, or the managed vault on the hosted plane), so a rogue "different tool call" cannot misuse a credential the agent never held.

Packages

`@marchward/engine`: the deterministic decision engine + hash-chain audit primitive (zero dependencies).
`@marchward/proxy`: the local governor (MCP + HTTP), --local or --remote.
`@marchward/sdk` (TypeScript) and `marchward` (Python, on PyPI): clients.

Project

Contributing: CONTRIBUTING.md (DCO sign-off, git commit -s)
Security: SECURITY.md · Conduct: CODE_OF_CONDUCT.md · Governance: GOVERNANCE.md
Trademark: TRADEMARK.md. The code is open (Apache-2.0); the name is not.
What's open vs. commercial: OPEN-VS-COMMERCIAL.md · our licensing commitment (the engine stays open, no relicensing rug-pull).
License: Apache-2.0.
Hosted free tier: https://app.marchward.ai/signup (500 decisions/mo, no card).