Like all of you I am impressed by the agentic harness both Claude Code and Codex CLI provide. At their core they are LLMs with a set of tools but we don't really know what's going on under the hood... So I built this to see all the underlying network traffic and parse it in real-time. — how many API calls per interaction, what the system prompts look like, token usage, subagent spawns, etc.
It's a local HTTP proxy + real-time dashboard. Point your AI agent at it with one env var and you see everything: requests, SSE streams, tool definitions, rate limits.
npm install -g watchtower-ai && watchtower-ai
And then go to your project and run your favorite CLI tool with the base URL set to the proxy.
Claude Code:
ANTHROPIC_BASE_URL=http://localhost:8024 claude
Some things I found interesting while building this: Claude Code sends 2-3 API calls per user message (quota check, token count, then the actual stream). It spawns subagents with completely different system prompts and smaller tool sets. The system prompt alone is 20k+ tokens.
This can be super useful if you also want to see the reasoning traces behind the scenes. IT is very rich information honestly and should enable you to build better agent harness.
Like all of you I am impressed by the agentic harness both Claude Code and Codex CLI provide. At their core they are LLMs with a set of tools but we don't really know what's going on under the hood... So I built this to see all the underlying network traffic and parse it in real-time. — how many API calls per interaction, what the system prompts look like, token usage, subagent spawns, etc.
It's a local HTTP proxy + real-time dashboard. Point your AI agent at it with one env var and you see everything: requests, SSE streams, tool definitions, rate limits.
npm install -g watchtower-ai && watchtower-ai
And then go to your project and run your favorite CLI tool with the base URL set to the proxy.
Claude Code: ANTHROPIC_BASE_URL=http://localhost:8024 claude
Codex CLI: OPENAI_BASE_URL=http://localhost:8024 codex
Some things I found interesting while building this: Claude Code sends 2-3 API calls per user message (quota check, token count, then the actual stream). It spawns subagents with completely different system prompts and smaller tool sets. The system prompt alone is 20k+ tokens.
This can be super useful if you also want to see the reasoning traces behind the scenes. IT is very rich information honestly and should enable you to build better agent harness.