jianshuo/ccglass

What it solves

Many AI coding agents (like Claude Code or Codex) use native apps or Node.js environments that ignore standard system proxy settings. This makes it nearly impossible to inspect the actual traffic—such as system prompts, tool schemas, and full message histories—sent to the LLM provider without using complex CA certificates or breaking the app via patching.

How it works

ccglass acts as a lightweight local reverse-proxy and web dashboard. It intercepts the plain HTTP hop between the coding agent and the AI API by redirecting the client's base-URL environment variable (e.g., ANTHROPIC_BASE_URL or OPENAI_BASE_URL) to localhost. For tools with hardcoded URLs, it provides a forward-proxy mode with TLS MITM to decrypt and log targeted traffic. All captured requests are stored in a content-addressed format to save space and can be viewed in real-time via a web interface.

Who it’s for

Developers using AI coding agents or IDE extensions (like Cursor, Cline, or Continue) who want to debug their agent's behavior, audit the exact prompts being sent, or track token usage and costs.

Highlights

• Live Request Stream: View full system prompts, tool calls, and message histories in real-time.
• Agent-Loop Visualization: A sequence diagram showing the flow between model tool selection and local execution results.
• Turn-to-Turn Diff: Compare two requests to see exactly what context was added or changed between turns.
• Cost & Token Tracking: Real-time calculation of input/output/cache tokens and estimated USD costs per request.
• Broad Compatibility: Supports a wide range of providers including Claude Code, DeepSeek-TUI, Ollama, OpenRouter, and various IDE extensions.
• Self-Inspection (MCP): When used with Claude Code, the agent can use MCP tools to inspect its own proxied requests within the chat.