See what your Claude Max subscription is really worth.
Quick Start · Features · How It Works · Contributing
demo.mp4
You're paying $200/month for Claude Max. You use it every day. But you have zero visibility into:
- What would this cost at API rates? Is Max actually saving you money?
- Which projects are eating the most tokens?
- What does your AI agent actually do all day?
- How much are you benefiting from prompt caching?
- When are your peak usage hours?
You're flying blind. AgentWatch fixes that.
AgentWatch reads your local Claude Code session logs and gives you a full observability dashboard — costs, tokens, tools, sessions, projects, and activity patterns — all running on your machine. No cloud. No accounts. No data leaves your computer.
Your Claude Max subscription: $200/month
What it would cost at API rates: $5,174/month
You save: $4,974/month (96%)
Not an enterprise platform. AgentWatch is a lightweight, local-first tool for individual developers. If you need multi-agent orchestration across teams, check out Paperclip.
git clone https://github.com/zmaxgong/agentwatch.git
cd agentwatch
# Start with your real Claude Code data
./start.sh --claude
# Or explore with demo data first
./start.shThen open dashboard/index.html in your browser. That's it.
# Watch for new Claude Code activity in real-time
./start.sh --watchNew events appear on the dashboard as you use Claude Code.
- Python 3.9+
- That's it. No Node.js, no Docker, no cloud accounts.
See exactly how much your Claude Max subscription saves you compared to API pricing — monthly projections, daily averages, per-model cost breakdowns, and prompt cache hit rates with estimated cache savings.
Hit "Copy as Image" to generate a stats card you can paste anywhere — Twitter, Slack, Discord. Shows your API value, monthly savings, top model, top project, total tokens, and session count.
Hour-of-day heatmap and day-of-week breakdown. See when you work with AI most, which hours cost the most, and what your agent actually does — categorized into Reading Code, Writing Code, Running Commands, and Thinking.
Click any session to replay it turn by turn. See every LLM call, every tool use, token counts, costs, and model choices — in chronological order.
| Tab | What it shows |
|---|---|
| Overview | KPIs, cumulative cost chart, model breakdown, per-project costs, period-over-period comparison |
| Cost & Savings | Shareable stats card, cache efficiency, savings calculator, 5h billing window, cost trends |
| Tools | Tool usage percentages at a glance, detailed breakdown bars, tools by project |
| Activity | Hour-of-day heatmap, day-of-week patterns, cost by time, "what your AI does" categories |
| Sessions | All sessions with sortable columns, click any row to replay turn-by-turn |
| Events | Full event log with type filtering and column sorting |
Claude Code logs AgentWatch Dashboard
~/.claude/projects/ Python backend Single HTML file
┌──────────────┐ ┌───────────────┐ ┌──────────────┐
│ JSONL session │ │ Import & │ │ Real-time │
│ logs (local) │────────>│ parse logs │───────>│ charts & │
│ │ │ SQLite store │ │ tables │
│ │ │ REST API │ │ (Chart.js) │
└──────────────┘ └───────────────┘ └──────────────┘
Everything stays on your machine. No cloud services. No telemetry. No data collection. The SQLite database lives in /tmp by default and is never transmitted anywhere.
AgentWatch is 100% local by design:
- Your Claude Code session logs are read from your local filesystem
- All data stays in a local SQLite database
- The dashboard is a static HTML file — no external server
- No analytics, no tracking, no phone-home
- The backend runs on
localhostonly - Nothing is ever uploaded, shared, or transmitted
When you open-source your fork or share screenshots, only the aggregated stats you explicitly copy are shared — never raw session data.
If you use the Anthropic API directly (not Claude Code), wrap your client:
from agentwatch import AgentWatch, MonitoredClient, AgentWatchConfig
aw = AgentWatch(AgentWatchConfig(project_id="my-app"))
aw.start_session()
# Wrap your Anthropic client — everything else stays the same
client = MonitoredClient(aw)
response = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=1024,
messages=[{"role": "user", "content": "Hello"}]
)Or wrap an existing client:
import anthropic
existing = anthropic.Anthropic(base_url="...", api_key="...")
client = MonitoredClient(aw, client=existing)agentwatch/
├── sdk/ # Python SDK (zero mandatory dependencies)
│ └── agentwatch/
│ ├── client.py # Core telemetry client
│ ├── wrapper.py # Drop-in Anthropic wrapper
│ ├── claude_code_import.py # Claude Code log importer
│ ├── config.py # Configuration & model pricing
│ ├── events.py # Event types & data structures
│ └── detectors.py # Hallucination, drift, security detection
├── backend/
│ └── server.py # FastAPI backend with SQLite
├── dashboard/
│ ├── index.html # Self-contained dashboard (vanilla JS + Chart.js)
│ └── assets/ # Logo, favicons
├── examples/
│ ├── quickstart.py # Anthropic wrapper example
│ ├── manual_tracking.py # Manual recording example
│ └── demo_data.py # Realistic demo data generator
├── tests/ # Test suite
├── start.sh # One-command setup
└── .github/workflows/ci.yml # CI pipeline
AgentWatch works out of the box. For customization:
| Environment Variable | Default | Description |
|---|---|---|
AGENTWATCH_DB |
/tmp/agentwatch.db |
SQLite database path |
SDK options (cost alert thresholds, detection toggles, batching) are in sdk/agentwatch/config.py.
See CONTRIBUTING.md. Bug reports, feature requests, and PRs are welcome.
See SECURITY.md for our security policy.
MIT — see LICENSE.
Built by Tandm Labs — a human-AI company.