看得见，才谈得上理解。我写了个 ccglass，npm 装上就能看到 Claude Code 背地里给大模型发了什么——系统提示词、48 个工具、token 账，全摊开。https://t.co/99thvxcoqh

jianshuo/ccglass

Source: https://github.com/jianshuo/ccglass

ccglass

See exactly what your coding agent sends to the model. A zero-dependency local logging reverse-proxy + web dashboard for Claude Code, Codex, DeepSeek-TUI, and Kimi. One command, like ollama:

npm install -g ccglass
ccglass

Run with no arguments and ccglass asks which client to inspect:

  Which client do you want to inspect?

    1) Claude Code
    2) Codex (OpenAI)
    3) DeepSeek-TUI
    4) Kimi (Moonshot, via Claude Code)

  >

Or name it directly: ccglass claude, ccglass codex, ccglass deepseek, ccglass deepseek-tui, or ccglass kimi.

ccglass starts a proxy, points the client at it via the right base-URL env var, launches it for you, and opens a dashboard where you watch every request in real time — the full system prompt, every tool schema, the message history, token/cache/cost numbers, and a turn-to-turn diff.

  ● ccglass watching Codex (OpenAI) → https://api.openai.com
    dashboard: http://127.0.0.1:57633

Why

These CLIs are Node/native apps that ignore HTTP_PROXY/HTTPS_PROXY — so Charles/mitmproxy never see the traffic, and fetch-patching tools break across updates. ccglass sidesteps all of it: the client does the HTTPS to the real API itself; you only intercept the plain HTTP hop to localhost. No CA certs, no TLS pinning.

Supported clients

Kimi runs through Claude Code against Moonshot’s Anthropic-compatible endpoint — make sure your Moonshot key is set (ANTHROPIC_AUTH_TOKEN). DeepSeek-TUI uses its OpenAI-compatible Chat Completions endpoint — make sure your DeepSeek key is set (DEEPSEEK_API_KEY).

What you get

• Live request stream — every call appears instantly; click to expand the system prompt, messages, and tools with all escaped strings unescaped.
• Turn-to-turn diff — pick two requests, see exactly what context was added this turn and which blocks carry a cache breakpoint.
• Token / cache / cost — exact input/output/cache tokens from the response usage, cache-hit rate, and estimated USD per request (per-provider pricing).
• Response reassembly + export — streamed SSE rebuilt into the final message (stop_reason, tool calls, usage), for both the Anthropic and OpenAI wire formats; export any request to Markdown / JSON / HAR.

Usage

ccglass                       # pick a client interactively
ccglass claude [args...]      # inspect Claude Code (args pass through, e.g. --resume)
ccglass codex  [args...]      # inspect Codex
ccglass deepseek [args...]    # inspect DeepSeek-TUI (dispatcher)
ccglass deepseek-tui [args...] # inspect DeepSeek-TUI runtime directly
ccglass kimi   [args...]      # inspect Kimi (via Claude Code)
ccglass run --provider openai -- <cmd...>   # inspect any client
ccglass view                  # re-open the dashboard over saved .ccglass/ logs
ccglass export <id> --format md|json|har

Options

Logs & secrets

Captures are written to ./.ccglass/<session>/NNNN.json. Auth tokens (authorization, x-api-key) are masked by default — pass --no-redact to keep them. Treat the log directory as sensitive regardless.

Requirements

Node ≥ 18. No runtime dependencies.

Acknowledgments

Heartfelt thanks to 庄表伟 (@zhuangbiaowei) for contributing first-class DeepSeek-TUI support (#1).

DeepSeek-TUI ships as a dual-binary coding agent — a deepseek dispatcher and a deepseek-tui runtime. 庄表伟 wired up both as native ccglass providers, pointing them at the proxy via DEEPSEEK_BASE_URL and reusing the existing OpenAI-compatible Chat Completions adapter, so every DeepSeek request now shows up in the dashboard with zero extra setup. The contribution also added them to the interactive picker, documented usage across the README, and shipped provider regression tests to keep it working. Thank you for making ccglass better for the whole DeepSeek community. 🙏

License

MIT