@jianshuo: Only when you can see it can you talk about understanding. I wrote ccglass; install it via npm and you can see what Claude Code is secretly sending to the large model—system prompts, 48 tools, token accounts—all laid out.
Summary
ccglass is a zero-dependency local logging reverse proxy and web dashboard that lets you see in real-time the system prompts, tools, and token usage sent by coding agents like Claude Code to the large model.
View Cached Full Text
Cached at: 05/23/26, 02:11 PM
– | any client | per provider | per provider | per provider | 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 ``bash 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 -- # inspect any client ccglass view # re-open the dashboard over saved .ccglass/ logs ccglass export --format md|json|har `` ### Options | Flag | Default | Meaning | |---|---|---| | –provider
| from command | Force format/env forrun (claude/codex/deepseek/kimi/openai) | | –upstream | per provider | Override the upstream API | |–port | auto | Dashboard port | |–proxy-port | auto | Proxy port | |–dir |./.ccglass| Where logs are stored | |–open| off | Open the dashboard in your browser | |–no-redact| off | Keep auth tokens unmasked in saved logs | ## Logs & secrets Captures are written to./.ccglass//NNNN.json. Auth tokens (authorization, x-api-key) are **masked by default** — pass –no-redactto keep them. Treat the log directory as sensitive regardless. ## Requirements Node ≥ 18. No runtime dependencies. ## Acknowledgments Heartfelt thanks to **庄表伟 (@zhuangbiaowei (https://github.com/zhuangbiaowei))** for contributing **first-class DeepSeek-TUI support** (#1 (https://github.com/jianshuo/ccglass/pull/1)). DeepSeek-TUI ships as a dual-binary coding agent — adeepseekdispatcher and adeepseek-tuiruntime. 庄表伟 wired up both as native ccglass providers, pointing them at the proxy viaDEEPSEEK_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
Similar Articles
@CycleDecoded: Ridiculous, guys—a wild trick just stormed GitHub trending, directly exploiting a bug in the large model billing system. This thing is called pxpipe (MIT license), a local open-source proxy tool specifically designed to counter Claude Code's billing shock. The principle is absolutely genius: large models charge text tokens by word count, but images are billed by fixed pixels. So it simply takes your long, bloated system prompts, code, and history logs, snaps them into a dense PNG image, and feeds it through the model's vision channel. This is a brutal move—a direct bypass of expensive text billing!
pxpipe is a local open-source proxy tool that reduces Claude Code bills by approximately 70% by rendering large amounts of text (such as system prompts, code, and logs) into PNG images and feeding them through the large model's vision channel. It exploits the fact that images are billed by pixel rather than by word count. The tool is perfectly compatible with the Fable 5 model, operates with clever automation, but uses lossy compression and is unsuitable for sensitive data.
@koffuxu: https://x.com/koffuxu/status/2054527573439181206
cc-switch is a desktop application that helps users switch the underlying large model used by Claude Code with one click through a visual interface, without manually modifying configuration files. The article introduces how to install and configure it to access the DeepSeek model, and explains how it works.
@interjc: I connected Baishan Zhisuan's models to Claude Code via cc-switch, basically achieving GLM5.2 freedom. Now just waiting for when Kimi K3 can be deployed.
The user connected Baishan Zhisuan's models to Claude Code via the cc-switch tool, achieving free use of GLM5.2, and looks forward to the subsequent deployment of Kimi K3.
@yiyangleex: Using tools like Codex and Cursor to learn and understand codebases has a major issue: if you want to deeply understand a project's architecture, they can often only provide you with a bunch of markdown text or Mermaid diagrams.
Points out that current AI code understanding tools like Codex/Cursor can only provide static text or Mermaid diagrams, while modus can generate interactive diagrams and sandbox HTML in conversation, allowing users to intuitively understand code architecture through dragging and clicking, demonstrated with the Codex CLI source code as an example.
@vintcessun: An 8-stage vulnerability discovery agent that runs on a Claude subscription, essentially a reproduction of the Cloudflare Project Glasswing paper. Multiple narrow agents + intentional disagreement verification + reachability gating, breaking down 'which piece of code can actually be exploited by an attacker' into 8 precise steps. Recon uses Opus to decompose tasks, H…
Introduces an 8-stage vulnerability discovery agent based on Claude subscription, reproducing the core ideas of the Cloudflare Project Glasswing paper. It uses multiple narrow agents, intentional disagreement verification, and reachability gating to break down exploit analysis into precise steps, without the need for API keys.