内容
# mitmproxy MCP Server
A Model Context Protocol (MCP) server that transforms mitmproxy into a powerful toolset for AI agents. This allows LLMs (like Claude, GPT-4, or local models) to inspect, modify, and replay HTTP/HTTPS traffic in real-time.
## New in this build (the fun stuff)
This project has grown beyond simple capture + replay. You can now:
* **Search traffic quickly** with filters by domain/method/body text.
* **Extract data from responses** using JSONPath or CSS selectors.
* **Use session variables** to grab tokens from one response and reuse them in later replays.
* **Run lightweight fuzzing** against query params or JSON body fields for quick DAST checks.
* **Reconstruct API patterns** and export an OpenAPI-style spec from captured traffic.
* **Detect auth patterns** (Bearer/JWT/API key/session cookie/CSRF/OAuth hints).
* **Generate starter scraper code** from real captured flows.
## Why use this?
Standard "web search" or "fetch" tools are stateless and easily detected. mitmproxy-mcp provides:
* **Deep Debugging**: The agent can inspect full request/response cycles (headers, payloads, cookies) to identify why a frontend is failing or why an API is returning a `4xx/500` error.
* **API Reverse Engineering**: Let the LLM observe undocumented internal APIs, map out JSON schemas, and generate client libraries or documentation automatically.
* **Automated Security Testing**: Perform DAST (Dynamic Application Security Testing) by allowing the agent to inject payloads into specific parameters and analyze the response.
* **Live Interception**: Modify traffic on the fly: inject headers, mock responses for testing, or block tracking pixels to reduce noise.
* **Stealth Replay**: Uses `curl-cffi` to mimic Chrome/Safari TLS fingerprints, bypassing basic anti-bot measures that standard Python libraries trigger.
## Key Features
- **Lifecycle Control**: Start and stop the mitmproxy instance directly from the LLM.
- **Deep Inspection**: Capture full request/response cycles, including headers, bodies, and timing.
- **Traffic Search & Filtering**: Scope by domain, then search flows by query/method so context stays clean.
- **Active Interception**: Dynamic rules to inject headers, replace body content via regex, or block requests.
- **Stealth Replay**: Re-execute flows using `curl-cffi` to impersonate modern browser TLS fingerprints (e.g., Chrome).
- **Session-Aware Workflows**: Extract dynamic values (like CSRF tokens) and re-inject them into replayed traffic.
- **Response Data Extraction**: Pull structured values from HTML/JSON responses via CSS selectors and JSONPath.
- **API Discovery**: Group similar endpoints, identify path/query params, and export OpenAPI JSON.
- **Security Recon Helpers**: Detect common auth patterns and run targeted fuzz payloads.
- **Scraper Bootstrapping**: Generate executable `curl-cffi` automation code from observed traffic.
## Quickstart
### Option 1: Using `uvx` (Recommended)
Add this to your MCP client configuration (e.g., Claude Desktop, Cursor, or AntiGravity):
```json
{
"mcpServers": {
"mitmproxy-mcp": {
"command": "uvx",
"args": ["mitmproxy-mcp"]
}
}
}
```
### Option 2: From a Cloned Repository
If you want to run from a local clone (useful for development or testing unreleased changes):
```bash
git clone https://github.com/snapspecter/mitmproxy-mcp.git
```
Then add this to your MCP client configuration, replacing the path with your clone location:
```json
{
"mcpServers": {
"mitmproxy-mcp": {
"type": "stdio",
"command": "uv",
"args": ["run", "--directory", "/path/to/mitmproxy-mcp", "mitmproxy-mcp"]
}
}
}
```
## Installation
### Option 1: Global Install (with `uv`)
```bash
uv tool install mitmproxy-mcp
```
### Option 2: Docker (Isolated Environment)
```bash
# Build and run
docker build -t mitmproxy-mcp .
docker run -p 8080:8080 mitmproxy-mcp
```
### Option 3: Manual Pip Install
```bash
python -m venv venv
source venv/bin/activate
pip install mitmproxy-mcp
```
## Available Tools
### Lifecycle & Configuration
* `start_proxy(port=8080)`: Starts the mitmproxy server.
* `stop_proxy()`: Shuts down the proxy.
* `set_scope(allowed_domains)`: Filters recorded traffic (e.g., `["api.github.com", "example.com"]`).
* `list_tools()`: Returns all exposed MCP tools and input schemas.
### Inspection
* `get_traffic_summary(limit=20)`: Returns a list of recent network flows.
* `inspect_flow(flow_id)`: Provides full details and a `curl` equivalent for a specific flow.
* `search_traffic(query, domain, method, limit=50)`: Filter captured traffic by keyword/domain/method.
* `clear_traffic()`: Clears persisted traffic history.
### Modification & Interception
* `add_interception_rule(rule_id, action_type, ...)`:
* `action_type`: `inject_header`, `replace_body`, or `block`.
* `phase`: `request` or `response`.
* `set_global_header(key, value)`: Injects a header into every request.
* `remove_global_header(key)`: Removes a previously injected global header.
* `list_rules()`: Shows currently active interception rules.
* `clear_rules()`: Flushes all active interception rules.
### Replay
* `replay_flow(flow_id, method, headers_json, body)`: Re-sends a request with modifications using browser-grade impersonation.
### Extraction & Session Variables
* `extract_from_flow(flow_id, json_path=None, css_selector=None)`: Extract values from captured response bodies.
* `set_session_variable(name, value)`: Manually set a session variable.
* `extract_session_variable(name, flow_id, regex_pattern, group_index=1)`: Extract and save a variable from a response body.
### Security & API Analysis
* `fuzz_endpoint(flow_id, target_param, param_type, payload_category, timeout=10.0)`: Replay with payload substitutions and report anomalies.
* `get_api_patterns(domain=None, limit=50)`: Cluster captured traffic into endpoint patterns.
* `export_openapi_spec(domain=None, limit=50)`: Produce an OpenAPI v3 JSON spec from captured traffic.
* `detect_auth_pattern(flow_ids=None)`: Infer likely auth mechanisms from observed requests.
### Code Generation
* `generate_scraper_code(flow_ids, target_framework="curl_cffi")`: Generate starter scraper/automation code from captured flows.
## Programmatic Usage
Note: These are JSON-RPC calls sent by the MCP Host (Client). You do not need to type these manually in the terminal.
1. **Initialize the Proxy**:
`{"method": "tools/call", "params": {"name": "start_proxy", "arguments": {"port": 8080}}}`
2. **Intercept & Block**:
`{"method": "tools/call", "params": {"name": "add_interception_rule", "arguments": {"rule_id": "block-ads", "action_type": "block", "url_pattern": ".*analytics.*"}}}`
3. **Modify Response**:
`{"method": "tools/call", "params": {"name": "add_interception_rule", "arguments": {"rule_id": "mock-api", "action_type": "replace_body", "url_pattern": ".*user/profile.*", "action_value": "{\"name\": \"AI Agent\"}"}}}`
## Helpful Tips
* **Manage Context**: Use `set_scope` immediately. LLMs perform poorly when flooded with background OS telemetry.
* **Browser Setup**: Ensure your browser or application is configured to use the proxy (usually `localhost:8080`) and has the mitmproxy CA certificates installed for HTTPS inspection.
* **Stealth**: The `replay_flow` tool uses `curl-cffi` specifically to avoid being flagged as a bot by services that check TLS fingerprints.
## Development
```bash
git clone [https://github.com/snapspecter/mitmproxy-mcp.git](https://github.com/snapspecter/mitmproxy-mcp.git)
cd mitmproxy-mcp
uv sync
uv run pytest
```
**License:** MIT
```
