AgentLens

Full-stack, zero-code observability for CLI-based AI Agents.

AgentLens is a passive observability platform designed to give you complete visibility into AI CLI agent behavior, directly from your terminal to a beautiful local dashboard.

Instead of requiring you to embed SDKs or modify your agent's source code, AgentLens operates as a Man-in-the-Middle (MITM) reverse proxy. It transparently intercepts API calls made by the agent CLI to the underlying LLM provider, extracting session data, tokens, latency, and tool executions.

AgentLens natively supports:

• Claude Code (Anthropic)
• Antigravity CLI / agy (Google Gemini)

Features

• Zero-Code Integration: Works by simply proxying HTTP traffic. No SDKs, no code changes.
• Native Claude Session Sync: Automatically extracts Claude Code's internal CLI session IDs to perfectly sync your terminal sessions with the dashboard.
• Hallucination Detection: Flags contradictions between the LLM's claims and the actual tool results.
• Frustration Analyzer: Scores user frustration based on behavioral (e.g. rage prompting) and linguistic signals using LLM-as-a-judge.
• Real-time Dashboard: A built-in React-style UI showing a complete timeline of every session.

Screenshots

Installation (One-Click)

AgentLens comes with an automated installer that will set up Docker Compose, generate your local SSL certificates for MITM, and install the lens CLI wrapper.

./install.sh

This will automatically spin up:

• AgentLens Gateway: http://localhost:8080
• AgentLens Dashboard: http://localhost:8090
• PostgreSQL: localhost:5432

Open your browser and navigate to http://localhost:8090 to view the dashboard!

Usage (lens wrapper)

Instead of manually exporting global proxy variables and polluting your terminal environment, AgentLens installs a lightweight wrapper command called lens.

Simply prepend lens to whatever agent command you want to run. It dynamically injects the proxy settings only for that specific execution, ensuring that all other background terminal traffic (like normal git or curl commands) remains completely unaffected!

lens claude
# or
lens agy
# or (alpha)
lens python my_agent.py

Architecture Overview

AgentLens intercepts traffic asynchronously to prevent adding latency to your terminal experience:

1. The Agent CLI sends an HTTP request to the Gateway via the lens proxy override.
2. The Gateway computes a deterministic session ID and forwards the request to the real LLM provider.
3. The response is buffered and sent back to the CLI immediately.
4. Asynchronously in the background, the Gateway parses the buffered payload, extracts token usage, tool calls, and text, and persists it to PostgreSQL.
5. A background Evaluator worker runs LLM-as-a-judge on idle sessions to detect hallucinations and user frustration.

Configuration

You can configure the upstream LLM endpoints by editing the config.yaml file located in the root directory. For example, to change your Gemini endpoint:

proxy:
  gemini_upstream: "https://your-custom-endpoint.com"

After modifying the file, simply run make deploy again to rebuild the image and apply the changes.