Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
<div align="center">
AgentLens
Chrome DevTools for AI Agents
See every decision, trace every tool call, catch every hallucination.
Quick Start · Features · VS Code Extension · GitHub Actions CI · Supported Frameworks · MCP Integration · Examples · Contributing
</div>
Why AgentLens?
AI agents fail in opaque ways — wrong data, hallucinated numbers, runaway costs, corrupted memory. AgentLens gives you a real-time debugging dashboard: every LLM call, tool execution, memory operation, and decision your agent makes — visible, searchable, replayable. Two lines of code to integrate, zero config to start.
Quick Start
1. Install and run the server
pip install agentlens-server
agentlens-server
# Server: http://localhost:8766 | Dashboard: http://localhost:51732. Install the Python SDK
pip install agentlens-sdk3. Add two lines to your agent
from agentlens_sdk import auto_instrument
auto_instrument()
# That's it. Open http://localhost:5173 and run your agent.TypeScript / Node.js
npm install @agentlens-sdk/sdkimport { autoInstrument } from '@agentlens-sdk/sdk';
autoInstrument();Run from source (development)
git clone https://github.com/ArkFelix7/agentlens
cd agentlens
make install && make devFeatures
Core Observability
| Feature | Description |
|---|---|
| Real-time Trace Graph | D3.js force-directed graph — every agent step as a node, click for full input/output |
| Cost Analytics | Per-model, per-step cost breakdown with cheaper-model suggestions |
| Hallucination Detection | Semantic comparison of tool outputs vs LLM responses, number transposition alerts |
| Memory Inspector | Version history, influence mapping, in-dashboard edit/delete for agent memory |
| Session Replay | VCR-style playback of any past run, shareable replay links |
v0.2 — New Features
| Feature | Description |
|---|---|
| Reliability Score Badge | 0–100 score (A/B/C/D) grading a session on hallucinations, errors, cost, and latency. Embeddable SVG badge for any README. |
| Budget Guardrails | Set per-session or per-model cost/token/call limits. Real-time alerts fire the moment a running agent crosses a threshold. |
| Auto Test Generation | One click turns any trace into a pytest fixture — captures inputs, outputs, and assertions so production failures become regression tests. |
| LLM Model Comparison | Replay any session with a different model and diff the outputs side-by-side. Compare cost, latency, and accuracy across GPT-4o, Claude, Gemini. |
| Prompt Version Control | Track every prompt edit, compare versions, and run A/B experiments across sessions — all without leaving the dashboard. |
| Multi-Agent Topology Map | Visual coordination graph for CrewAI, AutoGen, and custom multi-agent setups — see which agent called which, when, and at what cost. |
| Air-Gap Privacy Mode | Redact PII from traces before they reach the server. Full local-only mode with no external network calls. |
| VS Code / Cursor Extension | Inline cost and latency annotations on @trace decorated functions. Sidebar showing the last trace without leaving your editor. |
| GitHub Actions CI | Post a trace quality report as a PR comment — hallucination count, cost, reliability score, and test pass/fail. |
VS Code Extension
Install from the VS Code Marketplace or build from source:
cd vscode-extension
npm install
npm run compile
# Then: Extensions panel → "Install from VSIX..." → select the generated .vsixFeatures:
- Inline cost and latency annotations on any
@trace-decorated function - Sidebar panel with real-time trace feed
- Command palette:
AgentLens: Show Last Trace,AgentLens: Toggle Cost Annotations - Auto-connects to
http://localhost:8766on startup (configurable)
GitHub Actions CI
Add to any repo to get automatic trace quality checks on every PR:
# .github/workflows/agent-check.yml
name: Agent Trace Check
on: [pull_request]
jobs:
trace-check:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: ArkFelix7/agentlens/.github/actions/[email protected]
with:
script: python my_agent.py
fail-on-hallucination: 'true'
max-cost-usd: '0.10'Posts a comment to the PR with: reliability score, hallucination count, total cost, and latency breakdown.
Supported Frameworks
| Framework | Integration | How |
|---|---|---|
| OpenAI | Auto | auto_instrument() |
| Anthropic | Auto | auto_instrument() |
| LangChain | Callback | AgentLensCallbackHandler |
| CrewAI | Auto-detect | auto_instrument() |
| AutoGen | Auto-detect | auto_instrument() |
| Semantic Kernel | Filter | instrument_semantic_kernel(kernel) |
| Any Python | Decorator | @trace(name="my_step") |
| TypeScript/Node | Wrapper | trace(fn, { name: "my_step" }) |
| MCP agents | Zero-code | agentlens-mcp server |
MCP Integration
Zero-code observability for Claude Desktop, Cursor, Windsurf, and any MCP-compatible agent.
pip install agentlens-mcpAdd to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"agentlens": {
"command": "agentlens-mcp"
}
}
}Examples
See examples/ for five runnable demos:
demo_multi_step.py— full showcase: 8+ steps, intentional hallucination, memory ops (no API key needed)demo_multi_agent.py— multi-agent topology: orchestrator spawning researcher + writer agentsdemo_openai_agent.py— minimal@traceusagedemo_anthropic_agent.py—auto_instrument()usagedemo_langchain_agent.py— LangChain callback handler
make demoContributing
PRs welcome. See CONTRIBUTING.md for setup instructions and the architecture overview.
License
MIT