fastmcp-agents 0.1.2

Fastmcp Agents project

• Are you tired of teaching every Agent how to use every tool?
• Why put the instructions on how to run git_clone into every Agent you write?
• Why do you have to keep telling it that it cant clone with depth: 0?

FastMCP-Agents is a framework for building Agents into yours and more importantly, other people's MCP Servers!

How's it work?!

Just take any third-party MCP Server and add just one extra tool -- an embedded Agent that can use the tools the server provides!

Simply take your existing MCP Server

"mcp-server-tree-sitter": {
  "command": "uvx",
  "args": ["mcp-server-tree-sitter"]
}

And wrap it with an Agent:

"github_github-mcp-server": {
  "command": "uvx",
  "args": [
    "fastmcp_agents", "cli",
    "agent",
    "--name","ask_tree_sitter",
    "--description", "Ask the tree-sitter agent to find items in the codebase.",
    "--instructions", "You are a helpful assistant that provides users a simple way to find items in their codebase.",
    "wrap", "uvx", "mcp-server-tree-sitter"
  ]
}

There's more than just adding an AI Agent in FastMCP-Agents. You can also modify the tools and parameters of the server to make it easier for the Agent to use.

You can use FastMCP-Agents to wrap any MCP Server via the command line, configure the transformation with a YAML or JSON file, or even write Python code to configure the transformations!

Option	Agents	Servers	Override Tools	Wrap Tools
Python	∞	∞	Yes	Yes
YAML or JSON	∞	∞	Yes	No
Command-line	∞	1	No	No

Example Servers

Here are some example servers that you can use to get started. You can find the full list of bundled servers here.

Using FastMCP-Agents as a CLI or MCP Server

For all of the following options start with:

1. Install UV
2. Follow the instructions for configuring your preferred provider and model
3. Follow the instruction for your MCP Client (Web UI, IDE (VSCode, Roo Code), cli)

Providers

Google Gemini

1. Set up your Google Gemini credentials. gcloud init should be your first option.
2. export MODEL="gemini/gemini-2.5-flash-preview-05-20"

Alternatives to gcloud init:

1. Create a gemini api key
2. export GEMINI_API_KEY=your-gemini-api-key

Google Vertex AI

1. Set up your Google Vertex AI credentials. gcloud init should be your first option.
2. Set your model export MODEL="vertex_ai/gemini-2.5-flash-preview-05-20"

Alternatives to gcloud init:

1. Create a service account with Vertex AI User role
2. Download the credentials
3. Set the environment variable GOOGLE_APPLICATION_CREDENTIALS to the path of the JSON key file

MCP Clients

In each of these examples we'll use the wrale_mcp-server-tree-sitter as our MCP Server. Feel free to experiment with other MCP Servers.

CLI Tool Call Example

uvx fastmcp_agents cli \
agent \
--name "ask_tree_sitter" \
--description "Ask the tree-sitter agent to find items in the codebase." \
--instructions "You are a helpful assistant that provides users a simple way to find items in their codebase." \
call "ask_tree_sitter" "{\"instructions\": \"Analyze the codebase in . and tell me what you found.\"}" \
wrap uvx git+https://github.com/wrale/mcp-server-tree-sitter.git

MCP Inspector

1. Run npx @modelcontextprotocol/inspector uvx fastmcp_agents config --bundled wrale_mcp-server-tree-sitter
2. Visit http://localhost:6274/#tools
3. Click Connect to connect to your MCP Server
4. Click List Tools
5. Click ask_tree_sitter
6. Interact with the tool via the instructions text area

Open Webui

1. Run open-webui. This is the best way:

docker pull ghcr.io/open-webui/open-webui:main
docker rm -f open-webui
docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui -e WEBUI_AUTH=false --restart always ghcr.io/open-webui/open-webui:main

2. Run mcpo and your MCP Server to provide an OpenAPI interface for open webui to use: uvx mcpo --port 8000 -- uvx fastmcp_agents config --bundled wrale_mcp-server-tree-sitter
3. Visit http://127.0.0.1:3000
4. Register your tool with open webui. Click the account in the upper right and select settings > tools > (+) add connection. Set the base url to http://localhost:8000 and click save.