Agent Lens — AI skill for Claude Code

AgentLens

**Video walkthrough:** https://youtu.be/ZLOEF98tE9k **60-second overview:** https://youtu.be/s48Y7D9n-N4

Self-hosted observability + LLM-as-judge eval dashboard for multi-agent runs: live traces, call trees, tool I/O, and inline rubric scoring.


What it is

AgentLens makes multi-agent systems observable and evaluable without sending data to a third party. Wrap agent functions with a `@traced` decorator and every LLM call, tool invocation, agent handoff, and retry becomes a structured span. A React dashboard renders runs as an interactive Gantt timeline and hierarchical call tree, with per-span panels showing the exact prompt, output, tool call JSON, token counts, and latency.

On top of raw tracing, AgentLens ships a rubric-based eval harness. Define YAML rubrics describing what a good run looks like — "did the agent cite its sources?", "did it avoid redundant tool calls?" — and AgentLens runs Claude-as-judge against each completed trace. Pass/fail results with the judge's reasoning appear inline next to the trace that produced them.

Quickstart

git clone https://github.com/RitikPatill/agent-lens.git
cd agent-lens

# Install Python packages (requires Python 3.11+) and Node modules
py -3.11 -m pip install -e "sdk/[dev]"
py -3.11 -m pip install -e "server/[dev]"
cd dashboard && npm install && cd ..

# Set your API key, then start everything and run the built-in demo
export ANTHROPIC_API_KEY=sk-ant-...
make demo
# Open http://localhost:5173

`make demo` starts the FastAPI collector on port 8000 and the Vite dashboard on port 5173, runs `examples/research_assistant.py`, then keeps both services alive. Press Ctrl-C to stop.

Usage

After `make demo` the dashboard opens at `http://localhost:5173`. The demo run (a Planner → Researcher → Writer pipeline ans