Agent MCP Server

Agent MCP Server is a specialized Model Context Protocol (MCP) server that acts as a middleman orchestrator. Instead of merely exposing raw tools (like file system access or web search) to an MCP client, this server exposes a single, powerful agent tool.

By calling the agent tool, your primary MCP client can spawn and delegate complex, multi-step background tasks to specialized, configured AI agents (powered by OpenAI, Anthropic, or Ollama). These agents can dynamically connect to other downstream MCP servers, invoke their tools in parallel, wait for long-running tasks to finish, and ask for user permission before executing sensitive actions.

🌟 Key Features

• Agent Delegation: Exposes an agent tool that dynamically lists available configured agents, allowing a parent LLM to delegate work to specialized sub-agents.
• Downstream MCP Federation: Connects to other Stdio and HTTP MCP servers, passing their tools to your configured agents.
• Multi-Provider Support: Supports OpenAI, Anthropic, and Ollama compatible LLM backends.
• Human-in-the-Loop (Elicitation): Natively supports the MCP Elicitation capability. If an agent tries to call a tool, the server can pause and ask the user for approval (Approve Once, Approve Always, Deny Once, Deny Always). Also supports forwarding the elicitation prompts of downstream MCP servers to the client.
• Granular Tool Filtering: Configure default tool policies (Ask (default), Allow, Deny) and specify AutoApproveTools or AutoDenyTools using glob or regex patterns.
• Long-Running Tasks: Built-in support for the MCP Tasks extension. Agents can trigger asynchronous downstream tools, and the server will automatically poll until completion without blocking the client.
• Dual Hosting Modes: Can run as a standard stdio MCP server or an http MCP server.
• Native AOT: Pre-compiled native binaries mean incredibly fast startup times, low memory usage, and no runtime dependencies needed.

🏗️ Architecture

1. MCP Client connects to Agent MCP Server.
2. Client asks to run the agent tool (e.g., "Use the 'Coder' agent to refactor this project").
3. The Server spins up an LLM chat loop for the "Coder" agent.
4. The "Coder" agent has access to downstream MCP servers (e.g., a local filesystem MCP).
5. The agent acts autonomously, fetching files, reading, and writing, subject to your configured approval policies.
6. Once finished, the final result is returned to the original MCP client.

🚀 Getting Started

Because Agent MCP Server is compiled with Native AOT, there are no prerequisites or SDKs required to run it.

Installation

Download the latest standalone binary for your operating system from the releases page.

Running the Server

You can run the server via standard I/O (default) or HTTP. You must provide a configuration, for example you can use a file for configuration and specify it via the --config flag.

# Run as stdio (default)
agent-mcp --config config.ini
# this is equivalent to 'agent-mcp stdio --config config.ini'

# Run as http server
agent-mcp http --config config.ini

⚙️ Configuration

Agent MCP Server can be configured in a variety of ways. Below there is an example using an .ini file format.

# Configure your LLM providers here.
# Supported Types: openai, anthropic, ollama

[Providers:claude]
Type = anthropic
ApiKey = sk-y…[redacted]

# Configure the MCP servers your agents can use.

[Mcp:filesystem]
Command = npx
Args:0 = -y
Args:1 = @modelcontextprotocol/server-filesystem
Args:2 = /home/myuser/projects/myproject

# Define specialized agents exposed to the client.

[Agents:coder]
Description = Use this agent to read and modify local files.
SystemPrompt = You are a principal software engineer. Complete the task step by step.
Provider = claude
Model = claude-fable-5-1
# Connects this agent to the filesystem downstream MCP
Mcp:0 = filesystem

For more configuration options and formats, please refer to the Wiki.

🛡️ Tool Permissions & Human-in-the-Loop

By default, Agent MCP Server uses Elicitation to ask for user approval before executing any downstream tools. You can fully customize this behavior by setting a default policy (Ask, Allow, Deny) or by using glob or regex patterns to automatically approve safe actions or block specific tools entirely. Additionally, when prompted, you can choose Approve Always or Deny Always to temporarily configure a tool's permissions on the fly without needing to edit your configuration file.

Read how to configure tool filtering in the Wiki.

🛠️ Exposed MCP Tools

Once running, the server exposes a single tool to the connected client:

• agent:
  • agent (string, enum): The specific agent to trigger (dynamically populated from your configuration, e.g., "researcher", "coder").
  • prompt (string): The task for the agent to perform.

📜 License

This project is released under the MIT License.