Make Your AI Coding Agent Transparent - See What It Actually Sends to the Model

If you've been using AI coding agents like Claude Code or Codex, you know how powerful they can be. But they also feel like a black box. What's actually in that system prompt? How much context is being sent every turn? Where are all my tokens going?

I recently found a tool called ccglass that answers these questions beautifully, and I wanted to share my experience.

What is ccglass?
ccglass is a lightweight local tool that acts as a reverse proxy between your AI coding agent and the model API. It intercepts all requests and responses, logs them, and displays them in a really nice web dashboard.

ccglass Dashboard

Getting Started
Installation is simple (Node.js 18+ required):

npm install -g ccglass

Then just run it and pick your agent:

ccglass

Or specify directly:

ccglass claude # for Claude Code
ccglass codex # for Codex
ccglass deepseek # for DeepSeek-TUI

It will:

1. Start a local proxy server
2. Set the right environment variables automatically
3. Launch the agent for you
4. Open the web dashboard

That's it! No CA certificates to install, no complicated setup.

What You Can See

The dashboard shows you everything:

1. The Full System Prompt

This is probably the most interesting part. You get to see how different agents frame their instructions to the model. Claude Code's system
prompt is fascinating to read!

1. Complete Message History

See the full context being sent each turn, how it evolves, and what gets kept vs. dropped.

1. Tool Schemas and Calls

See what tool definitions the agent provides to the model, and what tool calls the model makes in response.

1. Token and Cost Breakdown

Know exactly how many tokens you're using, what's cached, and get cost estimates per request and per session.

Token Summary

1. Turn-by-Turn Diff View

Compare requests to see exactly what changed between turns. Super useful for debugging why an agent started behaving differently.

Supported Agents

The list is pretty extensive:

• Claude Code (including Bedrock and Vertex modes)
• Codex (OpenAI)
• DeepSeek-TUI and Reasonix
• Kimi (Moonshot)
• OpenCode
• Ollama
• OpenRouter
• GLM/Zhipu
• CodeBuddy (VS Code/JetBrains plugins)

IDE Support

If you use Cursor, Cline, Continue.dev, or similar IDEs that let you set a custom API base URL, you can use the proxy mode:

ccglass proxy --provider openai

Then just point your IDE's API base URL to the local proxy address it gives you.

Why I Like This

1. Learn from the best - See how production-grade agent systems are built
2. Debug effectively - Understand why your agent is making certain choices
3. Optimize costs - See where your tokens are actually going
4. Local only - All logs stay on your machine (default redacts auth tokens)
5. MIT licensed - Completely open source

Pro Tip: Export for Documentation

You can export any request to raw HTTP, Markdown, JSON, or HAR format:

ccglass export / --format md

Great for documentation, bug reports, or just saving interesting prompts.

Try It Out

If you're using any AI coding agent regularly, I highly recommend giving ccglass a try. It will change how you think about these tools.

Install it now:

npm install -g ccglass

Check out the project: github.com/jianshuo/ccglass

What's your favorite AI coding agent? Have you found any good tools for understanding them better? Let me know in the comments!