Monitor, inspect, and debug all API traffic between AI coding agents and their APIs — with a real-time web dashboard.
Like Chrome DevTools Network tab, but for Claude Code and Codex CLI.
AI coding agents make dozens of API calls per interaction — streaming messages, token counts, quota checks, subagent spawns — and you can't see any of it. Watchtower sits between your AI agent and its API, captures everything, and gives you a live dashboard to inspect it all.
Works with:
- Claude Code (Anthropic)
- Codex CLI (OpenAI)
- Any OpenAI-compatible client
What you see:
- Every request and response, fully decoded
- SSE stream events in real time
- Live cost & token tracker — cumulative session cost and token count in the header, plus per-request cost from model list prices
- Search & filter — full-text search across models, types, tools, and message content; hide
token_count/quota_checknoise; show errors only - Conversation grouping — fold each turn's utility calls and subagents under its main request
- System-prompt & turn diffs — see exactly what changed vs the previous request
- Replay & breakpoints — re-send a captured request (optionally edited), or pause requests in-flight to edit or drop them ("Burp Suite for Claude")
- Copy buttons on every JSON block, system prompt, and message
- Agent hierarchy tracking (main agent → subagents → utility calls)
- Token usage, rate limits, and timing
- System prompts, tool definitions, full message history
- Reveal hidden extended thinking — optionally rewrite requests so the API streams thinking summaries (see Reveal thinking)
- Dark and light mode
- All logged to disk as JSON for later analysis
Zero runtime dependencies — it's a single script plus a self-contained dashboard. No npm install, no build step.
npm install -g watchtower-ai
watchtower-aiOr clone and run directly — there are no dependencies to install:
git clone https://github.com/fahd09/watchtower.git
cd watchtower
node intercept.mjsThen point your AI agent at the proxy:
# Claude Code
ANTHROPIC_BASE_URL=http://localhost:8024 claude
# Codex CLI (API-key auth)
OPENAI_BASE_URL=http://localhost:8024 codexOpen http://localhost:8025 — that's it. Both providers share the same proxy port. Requests are auto-detected.
Codex with a ChatGPT login (subscription auth)?
OPENAI_BASE_URLis ignored in that mode — Codex talks straight tochatgpt.comand the proxy never sees it. See Codex + ChatGPT auth for the one-time profile setup.
When Codex is logged in with a ChatGPT account (auth_mode: chatgpt), it sends its Responses API traffic directly to chatgpt.com/backend-api/codex and ignores OPENAI_BASE_URL — so pointing that env var at the proxy does nothing. Instead, add an opt-in Codex profile that routes the model calls through Watchtower while keeping your ChatGPT login.
Save this as ~/.codex/watchtower.config.toml (also included in the repo as docs/codex-watchtower.toml):
model_provider = "watchtower"
[model_providers.watchtower]
name = "watchtower"
base_url = "http://localhost:8024" # match your Watchtower proxy port
wire_api = "responses"
requires_openai_auth = true # keep the ChatGPT OAuth token + account idThen run Codex through the proxy:
codex --profile watchtowerCodex still attaches your ChatGPT OAuth token (requires_openai_auth); Watchtower forwards the request to chatgpt.com and captures it. Your default config.toml and the API-key OPENAI_BASE_URL path are untouched — it's opt-in per run.
# Default ports: proxy=8024, dashboard=8025
node intercept.mjs
# Custom ports
node intercept.mjs 9000 9001| Tab | What it shows |
|---|---|
| Overview | Duration, model, status, token counts, estimated cost, TTFT/throughput, context-window gauge, rate limits |
| Messages | Full conversation history (user/assistant messages) |
| Turn Diff | What changed in the messages vs the previous request |
| Sys Diff | Line-level diff of the system prompt vs the previous request |
| Response | Pretty-printed response JSON |
| Tools | Tool definitions with searchable parameters and schemas |
| Stream | Raw SSE events with expandable payloads |
| Headers | Request and response headers |
| Rate Limits | Rate limit headers with visual progress bars |
| Raw | Complete request/response bodies as JSON |
The proxy automatically classifies each request:
Anthropic:
messages_stream— Streaming chat (the main interaction)messages— Non-streaming chattoken_count— Token countingquota_check— Quota/permission check (max_tokens=1)
OpenAI:
responses_stream/responses— Responses API (Codex CLI default)chat_stream/chat_completion— Chat Completions APImodels— Model listingembeddings— Embedding requests
Each request is tagged with an agent role:
- main — The primary agent (Claude Code's main agent has the
Agenttool; Codex requests with tools) - subagent — A spawned sub-agent (Explore, Plan, etc.)
- utility — Token counts, quota checks, and tool-less calls
The header bar shows the running token count and estimated cost for the whole session, computed from each model's list price. Every request also shows its own cost in the Overview and Response tabs. Cache reads (~0.1×) and cache writes (~1.25×) are priced separately from fresh input. Unknown models show — instead of a wrong number.
Replay — open any captured request, click Replay / edit request, tweak the body (system, messages, tools, parameters), and send it upstream reusing the original request's auth. The result is captured as a new request tagged with a ↻ badge. Great for reproducing issues or A/B-testing a prompt change.
Breakpoints — flip the Breakpoints toggle in the sidebar and chat requests pause before being forwarded. A modal shows the request body: edit it and Forward, or Drop it entirely (the client gets a 499). Like Burp Suite, but for AI agents. Off by default; if the dashboard disconnects, breakpoints auto-disable and any paused requests are released so nothing hangs.
Replays and edited/dropped requests use the real upstream auth, which Watchtower keeps in memory only — never written to the JSON logs.
Reveal hidden thinking
On recent models (Opus 4.8/4.7, Sonnet 5, Fable 5), Claude Code requests extended thinking with display: "omitted" — the model reasons, but the API returns only a cryptographic signature, not the thinking text. Watchtower shows an empty thinking block by default and explains why.
To capture the reasoning, start the proxy with request rewriting enabled:
WATCHTOWER_REVEAL_THINKING=1 node intercept.mjsThis transparently rewrites Anthropic requests from display: "omitted" to "summarized", so the API streams a readable summary of the reasoning — which then appears in the dashboard. Off by default, so the proxy stays byte-for-byte transparent unless you opt in.
All requests are saved to ./logs/ as numbered JSON files:
logs/0001_messages_stream_claude-opus-4-6.json
logs/0002_token_count_claude-haiku-4-5-20251001.json
Each file contains the complete request/response cycle: headers, body, SSE events, timing, and rate limits.
Claude Code ──HTTP──┐
├──> Watchtower (proxy) ──HTTPS──> api.anthropic.com
Codex CLI ──HTTP──┘ │ ──HTTPS──> api.openai.com / chatgpt.com
│
├── logs to disk
├── broadcasts via WebSocket
│
Dashboard (web UI)
- Your AI agent sends requests to the local proxy instead of directly to the API
- Watchtower auto-detects the provider and forwards to the correct upstream
- Responses (including SSE streams) are decoded, logged, and forwarded back
- The dashboard connects via WebSocket for real-time updates
- Node.js >= 22 (uses the built-in crypto/streams; no npm dependencies)
Recently shipped:
- ✅ Cost & token tracking
- ✅ Search & filter (sidebar) + copy buttons
- ✅ Conversation grouping (fold each turn's requests into one unit)
- ✅ System-prompt diff & turn diff (what changed vs the previous request)
- ✅ Request replay & breakpoints (edit/drop in-flight — Burp Suite for Claude)
- ✅ Zero-dependency (dropped the
wspackage) - ✅ Transparent gzip pass-through (forwards original bytes to the client)
- ✅ Reveal hidden extended thinking (request rewriting)
- ✅ Observability: TTFT, throughput (tok/s), context-window gauge
Planned:
- Agent hierarchy tree visualization
- Export (session/request as JSON, HAR, or HTML)
Contributions welcome. Open an issue first for anything non-trivial.
git clone https://github.com/fahd09/watchtower.git
cd watchtower
node intercept.mjsThe proxy is in intercept.mjs, the zero-dependency WebSocket server in ws-server.mjs, provider logic in providers/, and the dashboard is a single dashboard.html. No dependencies, no build step, no bundler.