Claude Code Proxy

A tool to capture and inspect Claude Code API requests, including system prompts, tools, and message context.

Prerequisites

pip install mitmproxy

Run mitmdump --version once to generate the CA certificate in ~/.mitmproxy/.

Project Structure

claude_proxy/
├── run.sh          # Main runner script
├── env.sh          # Environment setup (source this)
├── README.md       # This file
├── src/
│   ├── capture.py  # mitmproxy capture script
│   └── export.py   # Export to readable text
└── output/         # Captured data goes here
    ├── claude_code_captured.json   # Latest capture (raw)
    ├── claude_code_log.jsonl       # History log
    └── capture_*.txt               # Exported readable files

Quick Start

Terminal 1 - Start the proxy:

./run.sh proxy

Terminal 2 - Run Claude through the proxy:

source env.sh
claude "hello"

# ...or pin a specific model to capture that model's prompt + tools:
claude --model claude-fable-5 -p "say hi"      # Fable 5 (headless)
claude --model claude-opus-4-8  -p "say hi"    # Opus 4.8 (headless)

-p runs Claude Code in headless/print mode — non-interactive and easy to script. Interactive claude captures a richer tool set; headless mode sends a smaller deferred tool set plus ToolSearch.

Terminal 2 - View the captured data:

./run.sh summary    # Quick summary
./run.sh export     # Export to readable .txt
./run.sh view       # Export and open in editor

Commands

Command	Description
./run.sh proxy	Start the mitmproxy server
./run.sh setup	Print environment variables to set
./run.sh summary	Show summary of captured data
./run.sh export	Export to readable text file
./run.sh view	Export and open in text editor
./run.sh clean	Remove all captured data
./run.sh help	Show help message

Captured Prompts

Ready-to-read captures live in per-model folders:

Folder	Model	Contents
fable/	claude-fable-5	System prompt, tools (readable + raw), full request

Each folder has its own README with capture details. Point Claude Code at a different model with --model <id> and re-run the proxy to add more.

What Gets Captured

• Model: The Claude model being used
• System Prompt: Full system instructions sent to Claude
• Tools: All available tools with their JSON schemas
• Message Count: Number of messages in the conversation context
• Max Tokens: Token limit for the response

Output Files

File	Description
output/claude_code_captured.json	Raw JSON of the latest capture
output/claude_code_log.jsonl	Append-only log of all captures
output/capture_*.txt	Human-readable exports

Manual Usage

If you prefer not to use the runner script:

# Terminal 1: Start proxy
mitmdump -s src/capture.py --listen-port 8080

# Terminal 2: Set environment and run Claude
export HTTPS_PROXY=http://127.0.0.1:8080
export HTTP_PROXY=http://127.0.0.1:8080
export NODE_EXTRA_CA_CERTS=~/.mitmproxy/mitmproxy-ca-cert.pem
claude "your prompt"

# View results
python3 src/export.py --summary
python3 src/export.py

How It Works

1. mitmproxy acts as a man-in-the-middle proxy
2. Claude Code is configured to route traffic through the proxy
3. The NODE_EXTRA_CA_CERTS environment variable makes Node.js trust the proxy's certificate
4. The capture.py script intercepts requests to api.anthropic.com and extracts relevant data
5. Data is saved to JSON and can be exported to human-readable format

Troubleshooting

"Certificate error" or "Connection refused"

• Make sure the proxy is running (./run.sh proxy)
• Ensure ~/.mitmproxy/mitmproxy-ca-cert.pem exists
• Re-run mitmdump --version to regenerate certificates

"No capture found"

• Make sure you ran Claude while the proxy was active
• Check that environment variables are set: echo $HTTPS_PROXY

Port 8080 already in use

• Find what's using it: lsof -i :8080
• Kill it or use a different port in src/capture.py and env.sh