dsh-mcp-bridge

Connect the entire MCP ecosystem to DeepSeek Harness with one file. Zero dependencies. No SDK.

Your DSH agent gets every tool from any MCP server — filesystem, GitHub, databases, search, memory, and thousands of community servers — as native tools it can call directly:

You: Create a GitHub PR for the changes in ./app
Agent: [calls mcp_github_create_pull_request] ✓ PR #42 created

✨ Why people use it

🚀 Quick start

A. dsh plugin add (recommended — the repo is a dsh.bundle)

dsh plugin --profile web add "github:Chhlafiu4312/dsh-mcp-bridge#main"

Then restart dsh web and configure servers via DSH_MCP_SERVERS (see below).

B. Dynamic plugin (one session, no files)

Open a DSH session, ask your agent to run the bridge via cordis_define (paste the code from plugin.js), or define it yourself — then talk to it:

// code.host body — minimal form
return {
inject: ['subprocess', 'tools', 'timer'],
apply(ctx) {
// ... paste the body of plugin.js here ...
},
}

C. Permanent (agent preset)

# 1. create your preset directory
mkdir -p ~/.dsh/.agent-presets/mcp-agent

# 2. copy the repo files in
cp plugin.js ~/.dsh/.agent-presets/mcp-agent/
cp agent.cordis.yml.example ~/.dsh/.agent-presets/mcp-agent/agent.cordis.yml
cp preset.yml.example ~/.dsh/.agent-presets/mcp-agent/preset.yml

# 3. (optional) configure your servers
export DSH_MCP_SERVERS='[
{"id":"filesystem","command":"npx","args":["-y","@modelcontextprotocol/server-filesystem","~/projects"]},
{"id":"github","command":"npx","args":["-y","@modelcontextprotocol/server-github"],"env":{"GITHUB_PERSONAL_ACCESS_TOKEN":"your_github_token"}}
]'

# 4. start a session on the `mcp-agent` preset — done

The agent instantly sees tools like mcp_filesystem_list_directory, mcp_github_create_issue, … plus mcp_status to inspect every connection.

🔧 Configuration

Servers are a JSON array. Each entry is either stdio or url:

{
"id": "my-server",                // tool prefix: mcp__
"command": "npx",                 // executable (stdio only)
"args": ["-y", "server-pkg"],     // argv after the executable
"cwd": "/abs/path",               // optional working directory
"env": { "KEY": "value" },        // optional environment
"url": "https://x.example/mcp",   // or: remote HTTP/SSE via mcp-remote
"disabled": false,                // keep configured but inactive
"allowTools": ["search_*"],       // optional: register only these tools
"denyTools": ["delete_*"],        // optional: hide these tools
"timeoutMs": 120000               // optional per-server call timeout
}

Set the array through DSH_MCP_SERVERS (JSON), or edit the defaults in resolveServers() inside plugin.js. Ready-made examples: examples/servers.json — filesystem, GitHub, fetch, memory, Postgres, remote.

🧠 How it works

Detailed walkthrough in docs/ARCHITECTURE.md:

DSH tools registry                    MCP server (child process)
─────────────────                    ───────────────────────────
mcp__  ── tools/call ──▶  npx         (stdio)
◀── text/JSON ────  └─ or mcp-remote  (HTTP)

1. Spawn — each configured server starts as a managed child process.
2. Handshake — initialize + notifications/initialized over newline-delimited JSON-RPC 2.0.
3. Discover — tools/list returns every tool with its JSON Schema.
4. Register — each tool is registered on the DSH tools registry; schemas are sanitized into the DSH parameter DSL.
5. Call — the model calls the tool; the bridge forwards tools/call and renders the text content back (120 s cooperative timeout, concurrency-safe).
6. Teardown — stopping the plugin unregisters the tools and terminates every process.

🛣️ Roadmap

• [ ] Auto-reconnect with backoff
• [ ] Native streamable-HTTP client (drop the mcp-remote dependency for url servers)
• [ ] Per-server permission prompts for destructive tools

🔒 Security notes

• MCP servers run with your permissions — only connect servers you trust.
• The filesystem server is scoped: give it only the directories the agent may touch.
• API tokens travel through the env field; prefer a secret manager or environment variables in production.
• Full trust model and hardening advice: SECURITY.md.

❓ FAQ

Can I use my own MCP server? Yes — anything that speaks MCP over stdio or streamable HTTP works. Point command/args at it (or url for remote servers) and its tools appear as mcp__.

What happens when a server crashes? The bridge retries up to 3 times with 2s/5s/10s backoff and re-registers the tools on the new connection. Check mcp_status for the current state.

Do I need to restart DSH after changing config? For the dsh plugin/preset install, yes — the config is read at plugin start. disabled: true keeps an entry configured without connecting it.

Why not the official MCP SDK? The SDK would drag a dependency tree into a plugin that only needs newline-delimited JSON-RPC. Zero dependencies keeps installs instant and reviewable (~400 lines).

Does it support resources and prompts? Yes — servers that declare those capabilities get mcp__list_resources, mcp__read_resource, mcp__list_prompts, and mcp__get_prompt tools automatically.

Upgrading from a pre-0.7.0 install? The npm package was renamed to dsh-mcp-connect — remove the old dependency once so the two layers never both activate: dsh plugin --profile web remove dsh-mcp-bridge.

npx-based servers fail with EPERM / "root-owned files"? Your npm cache was polluted by an old sudo npm run. Either fix it once:

sudo chown -R "$(id -u):$(id -g)" ~/.npm

or sidestep it per server by giving npx a writable cache:

{ "id": "filesystem", "command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/allowed/path"],
"env": { "npm_config_cache": "/path/to/writable/npm-cache" } }

🤝 Contributing

Issues and PRs welcome. node scripts/demo.mjs prints a one-shot demo transcript against the local fixture — the raw material for a README GIF. Test fixture included: test/mcp-http-test-server.cjs is a minimal streamable-HTTP MCP server you can run locally to verify the HTTP path end-to-end.

📄 License

MIT — see LICENSE.

dsh-mcp-bridge（中文）

一个文件，把整个 MCP 生态接进 DeepSeek Harness。零依赖，不用 SDK。

你的 DSH agent 可以直接调用任何 MCP server 的全部工具——文件系统、GitHub、数据库、搜索、记忆，以及社区里成千上万的 server：

你：把 ./app 的改动提一个 GitHub PR
Agent：[调用 mcp_github_create_pull_request] ✓ 已创建 PR #42

✨ 亮点

• 🔌 一行配置接入 1000+ 现成工具（MCP 生态是 agent 最大的工具目录）
• 🪶 零依赖：原生 JSON-RPC 2.0 over stdio，约 400 行纯 JavaScript
• 🌐 同时支持 stdio（本地 command）与 HTTP/SSE（远程 url，经 mcp-remote 桥接）
• 🔁 工具自动注册为 mcp__
• 🧹 插件停止/更新时自动清理子进程
• 🛡️ 不支持的 JSON-Schema 关键字安全降级，MCP server 始终是最终校验方

🚀 快速开始

A. 动态插件（单会话，不改文件）：在 DSH 会话中用 cordis_define 粘贴 plugin.js 的代码并运行。

B. 永久挂载（agent preset）：

mkdir -p ~/.dsh/.agent-presets/mcp-agent
cp plugin.js ~/.dsh/.agent-presets/mcp-agent/
cp agent.cordis.yml.example ~/.dsh/.agent-presets/mcp-agent/agent.cordis.yml
cp preset.yml.example ~/.dsh/.agent-presets/mcp-agent/preset.yml
export DSH_MCP_SERVERS='[{"id":"filesystem","command":"npx","args":["-y","@modelcontextprotocol/server-filesystem","~/projects"]}]'

然后用 mcp-agent 预设开新会话即可。工具命名规则：mcp__，另有 mcp_status 查看每个 server 的连接状态。

🔧 配置

server 配置是一个 JSON 数组，每项支持 id / command / args / cwd / env（stdio）或 url（HTTP，经 mcp-remote 桥接）。通过 DSH_MCP_SERVERS 环境变量或编辑 plugin.js 里的 DEFAULT_SERVERS 传入。示例见 examples/servers.json。

🧠 工作原理

spawn 子进程 → initialize 握手 → tools/list 发现工具 → 消毒 JSON Schema 并注册为 DSH 工具 → 模型调用时转发 tools/call → 渲染文本结果（120 秒协作超时、并发安全）→ 插件停止时注销工具并终止进程。

🛣️ 路线图 / 🔒 安全提示

路线图：resources/prompts 支持、断线自动重连、原生 streamable-HTTP 客户端、危险工具的逐服务器授权确认。安全：MCP server 以你的权限运行，只连接可信 server；filesystem server 只授予必要的目录；API token 走 env 字段，生产环境请用密钥管理器。

📄 License

MIT — 见 LICENSE。