MCP Conductor

MCPAI代理

提供方 niradler

AI代理 未标注 API Key 社区 / 第三方

conductor是一个AI代理工具网关,通过单一MCP端点聚合多个上游工具,提供实时认证、审计和访问控制。

一句话选型结论

按资料字段整理

MCP Conductor 属于AI代理类 MCP 服务器,传输方式未标注,API Key,没有现成安装命令,需按官方文档接入。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类AI代理 传输方式未标注 鉴权API Key 部署方式未标注
GitHub 星标 0 仓库热度,无公开仓库时显示未标注
工具数 1 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

AI代理
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量1暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

适用场景

AI代理TypeScript本地部署工具聚合访问控制实时审计MCP协议

安装与配置

未标注 API Key
上游未给出传输方式
  1. 1
    准备运行环境

    按官方文档确认运行环境(常见为 Node.js 或 Python),装好后再执行安装命令。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后客户端会列出约 1 个工具,随便调一个只读工具确认能返回结果。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
给代理接外部能力

让代理在任务中途调用搜索、数据库等真实工具。

多步任务编排

把一串操作拆成步骤,由代理按顺序执行。

自动化重复流程

把固定流程交给代理跑,减少手工操作。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 资料标注 1 个工具,职责相对集中,接入后不容易误操作。

先想清楚的地方

  • 没有现成安装命令,需要按官方文档自己填启动命令和参数。
  • 使用前要准备好API Key对应的凭据,并按最小权限配置。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在AI代理场景里让模型直接调用外部能力的人。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 不方便管理密钥或第三方授权的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
MCP Conductor当前 1 未标注 未标注 无 2026-05-17
Filesystem 560c4363 13 未标注 本地部署 无 2026-05-17
Server Everything (everything) 11 stdio · 本地进程 本地部署 有 2026-05-17
Time 7365cb6d 2 SSE · 服务端推送 远程部署 有 2026-05-18
Everything Claude Code 未标注 stdio · 本地进程 本地部署 有 2026-05-17
Openmemory 557e1585 未标注 stdio · 本地进程 本地部署 有 2026-05-17

服务器档案

AI代理
字段来自项目仓库与公开索引,中文说明由本站标注
提供方niradler上游标注的维护方 niradler
分类AI代理本站按上游分类展示
连接方式未标注上游未给出传输方式
鉴权方式API Key远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · 未标注本地安装时用到的包标识与运行时
部署类型未标注上游给出的部署形态
最近核验2026-05-17上游最后核验时间

详细介绍

12.3 KB
来自项目仓库与上游目录的原文,未做删改

conductor

A tool gateway for AI agents — one MCP endpoint, many upstreams, real auth, real audit.

conductor speaks Model Context Protocol to your agent and aggregates tools from any number of backends behind bearer-token auth, group-based access control, and a full call-by-call audit log. Agents stop caring how a tool is implemented; you stop pasting API keys into agent configs.

Backends implement a common ToolProvider interface. Today: MCP upstream servers over stdio. Planned: OpenAPI, GraphQL, HTTP tools, and sandboxed code & CLI execution via OpenShell — install your CLI in the sandbox image and call it through sandbox_exec instead of writing a one-off MCP wrapper.

Why this exists

Connecting many MCP servers directly to an agent is a context-tax trap. Each upstream's tool list and schemas get pre-loaded into the LLM's context window before the user has typed a word — the official GitHub MCP alone is ~50K tokens, and model accuracy starts dropping past ~100K ("Lost in the Middle"). Conductor sits in front of N upstreams as a single MCP endpoint and gives agents lazy, governed discovery: see only what the caller's role permits, list providers and tools on demand instead of all-at-once, audit and rate-limit every call, and (planned) pre-filter tools by intent so the model never sees the other 95%. The CLI-vs-MCP debate misses the point — use a local CLI when you want context-cheap and personal; use conductor when you need governance, audit, RBAC, and one endpoint across teams.```
┌──────────────┐ MCP/HTTP ┌───────────────────────────┐ stdio/… ┌─────────────┐
│ MCP client │ ──────────▶ │ conductor │ ──────────▶ │ upstream │
│ (agent/IDE) │ │ auth · groups · audit │ │ MCP server │
└──────────────┘ │ tool namespacing (__) │ └─────────────┘
└───────────────────────────┘ also: OpenAPI · GraphQL · …


## Highlights

- **One endpoint, many tools.** Upstreams are abstracted behind a `ToolProvider` contract; tools are exposed namespaced as `__`.
- **Lazy tool discovery.** Built-in `conductor__list_providers` and `conductor__list_tools` meta-tools let agents discover capabilities on demand instead of pre-loading every tool's schema into context. Provider-level descriptions and instructions are surfaced from each upstream's `serverInfo` / `initialize` payload.
- **Real auth.** SHA-256 hashed API keys, timing-safe comparison. Plaintext keys never appear in config.
- **Group-based access control.** Users belong to groups; groups grant providers (or `"*"` for all).
- **Audit everything.** Every tool call records user, provider, tool, redacted args, duration, status, and request id.
- **Observability built in.** Structured JSON logs; OpenTelemetry spans per call when `OTEL_EXPORTER_OTLP_ENDPOINT` is set.
- **Graceful lifecycle.** Clean LIFO shutdown: HTTP → sessions → providers → audit → telemetry.
- **Zero-ceremony start.** Ships a working config with a dummy key — clone, install, run, connect.

## Quick start

> [!NOTE]
> Requires **Node.js ≥ 20.11** and **pnpm ≥ 9**.```bash
pnpm install
pnpm build
CONDUCTOR_CONFIG=examples/conductor.json pnpm dev

The bundled examples/conductor.json boots with user alice (API key changeme, hash baked in) and one upstream provider (everything via npx @modelcontextprotocol/server-everything).

Connect any MCP client:```
URL: http://127.0.0.1:18080/mcp
Header: Authorization: Bearer changeme


13 tools appear, namespaced as `everything__*`. Sanity check:```bash
curl http://127.0.0.1:18080/health
# {"ok":true,"sessions":0,"providers":["everything"]}

Use your own API key```bash

pnpm hash-key my-secret-key

sha256:7f83b1657ff1fc53b92dc18148a1d65dfc2d4b1fa3d677284addd200126d9069


Copy `examples/conductor.json`, paste the hash into `users[].apiKeyHash`, point `CONDUCTOR_CONFIG` at your copy.

## Configuration

A single JSON file, validated by Zod at load time. Schema lives at `packages/server/src/conductor-config.ts`. Unknown keys are rejected.```jsonc
{
  "server": { "host": "127.0.0.1", "port": 18080, "maxSessions": 100 },
  "users": [
    { "name": "alice", "apiKeyHash": "sha256:…", "groups": ["admins"] }
  ],
  "groups": [{ "name": "admins", "providers": ["*"] }],
  "providers": [
    {
      "type": "mcp",
      "name": "everything",
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-everything"],
      "env": {}
    }
  ],
  "audit": { "type": "console" },
  "telemetry": { "serviceName": "conductor", "otlpEndpoint": "" }
}

Top-level fields

Field Purpose
server host, port, maxSessions — bind address and session LRU cap.
users[] name, apiKeyHash (sha256:<64 hex>), groups[].
groups[] name, providers[] (provider names, or ["*"] for all).
providers[] Discriminated on type. Today: "mcp" (stdio). More types incoming.
audit Currently { "type": "console" } with optional bufferSize.
telemetry serviceName + otlpEndpoint. Traces are a no-op if the endpoint is unset.

MCP provider options

Field Default Notes
command, args, env — How to spawn the upstream.
initialListTimeoutMs 15000 Deadline for the first tools/list after connect.
callTimeoutMs 60000 Per-call timeout forwarded as an AbortSignal.
reconnect.maxAttempts 10 Exponential backoff between initialDelayMs and maxDelayMs.
reconnect.initialDelayMs 1000
reconnect.maxDelayMs 30000

Environment variables

Var Meaning Default
CONDUCTOR_CONFIG Path to conductor.json. ./conductor.json
PORT Overrides server.port after config load. —
LOG_LEVEL debug, info, warn, error. info
OTEL_EXPORTER_OTLP_ENDPOINT OTLP HTTP base URL. Traces disabled if unset. —

> [!IMPORTANT]
> Never commit plaintext API keys. Run pnpm hash-key and store only the resulting sha256:… digest.

Architecture

conductor is a pnpm workspace with strict dependency layering — core has no sibling deps, providers and the gateway depend only on core, and server wires it all together.

Package Role
@mcp-conductor/core ToolProvider interface, ProviderRegistry, audit & config stores, logger, OTel bootstrap, shutdown registry. No sibling deps.
@mcp-conductor/provider-mcp ToolProvider that spawns an upstream MCP server over stdio. Handles timeouts and reconnect.
@mcp-conductor/provider-openshell ToolProvider for NVIDIA OpenShell (gRPC). Protos vendored; regenerate with pnpm update-openshell-protos. Currently a stub.
@mcp-conductor/gateway The HTTP MCP server. Auth, groups, audit wrapping, namespacing, session manager. Accepts any ToolProvider[].
@mcp-conductor/server The CLI (conductor). Loads conductor.json, wires providers into the gateway, owns signals.

Request lifecycle

  1. Authenticate. Bearer token → timing-safe SHA-256 compare against each user's apiKeyHash. Bad token → 401.
  2. Resolve access. User's groups → set of reachable providers ("*" expands to all configured).
  3. Session. New session = MCP initialize request. The gateway assembles a per-session MCP server that only advertises tools the caller may see.
  4. Namespace. Tools are registered as __. The separator __ is reserved; provider names must not contain it.
  5. Call. The handler decodes __, forwards to the audit-wrapped provider with { user, requestId, signal }. Providers must honour signal for cancellation.
  6. Audit + trace. One AuditStore.insertCall per invocation with redacted args, status, duration, request id. One OTel span per call.
  7. Shutdown. LIFO: HTTP server → sessions → providers → audit → telemetry.

Repository layout```

packages/
core/ foundation: ToolProvider, stores, logger, OTel, lifecycle
provider-mcp/ stdio MCP upstream adapter
provider-openshell/ OpenShell gRPC provider (stub; protos vendored)
gateway/ HTTP MCP server, auth, groups, audit, namespacing
server/ CLI binary: config + wiring
examples/
conductor.json working example config (alice / changeme)
scripts/
hash-api-key.ts prints sha256:<hex> for an API key
docs/
plans/ roadmap and implementation plans


## Writing a provider

Implement `ToolProvider` from `@mcp-conductor/core`:```ts
import type {
  ToolProvider,
  ToolSpec,
  ToolCallContext,
  ToolCallResult,
} from &quot;@mcp-conductor/core&quot;;

export class MyProvider implements ToolProvider {
  readonly name = &quot;my-provider&quot;;

  async connect(): Promise&lt;void&gt; { /* … */ }
  async close(): Promise&lt;void&gt; { /* … */ }
  async listTools(): Promise&lt;ToolSpec[]&gt; { /* … */ }

  async callTool(
    name: string,
    args: unknown,
    ctx: ToolCallContext,
  ): Promise&lt;ToolCallResult&gt; {
    // honour ctx.signal for cancellation
  }
}

Register it with the ProviderRegistry — or, for config-driven loading, add a branch to packages/server/src/provider-factory.ts and extend ProviderEntrySchema.

Development```bash

pnpm build # tsc -p tsconfig.build.json per package
pnpm typecheck # tsc --noEmit (includes tests)
pnpm test # vitest — 255 tests across all packages
pnpm test:watch
pnpm clean # rm -rf dist/ .tsbuildinfo


Run a single test file or test name:```bash
pnpm test -- packages/gateway/tests/auth.test.ts
pnpm test -- packages/gateway/tests/auth.test.ts -t &quot;rejects invalid bearer&quot;

> [!IMPORTANT]
> This is ESM-only (&quot;type&quot;: &quot;module&quot;, NodeNext resolution). Local imports must use the .js extension — TypeScript compiles to ESM .js and NodeNext will not resolve extensionless imports.

Protocol notes

  • Transport. Streamable HTTP (@modelcontextprotocol/sdk). One initialize request opens a session; subsequent requests must carry mcp-session-id. Non-initialize traffic without a session gets 400.
  • Tool schema. The gateway passes upstream JSON Schema through as z.unknown() per property. Upstream remains the source of truth for validation — the gateway never rewrites tool inputs.
  • Correlation. Every request gets an X-Request-Id (echoed on the response and threaded through ToolCallContext and the audit log).

> [!TIP]
> See docs/plans/ for the roadmap — OpenAPI, GraphQL, HTTP tools, and OpenShell sandboxed execution are next.

常见问题

5 问
答案由本站依据条目字段整理,不引用上游原文
MCP Conductor 怎么安装?

资料没有给出安装命令,请到官方仓库或文档按说明接入;这类条目多数是生态应用,或尚未发布可安装包。

MCP Conductor 怎么在客户端里配置?

它的传输方式未标注,在客户端 MCP 配置里按官方文档填写启动命令和参数,保存后重启客户端生效。

MCP Conductor 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

MCP Conductor 需要 API Key 吗?

资料标注需要API Key认证,使用前准备好对应凭据,并按最小权限配置。

MCP Conductor 能提供哪些能力?

资料显示约 1 个工具;能力标签包括 AI代理、访问控制、TypeScript。具体工具清单以官方文档为准。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 2026-05-17。