AgentLens

Local-first diagnostics & observability for AI agents. One line to integrate, a single SQLite file, a built-in dashboard. AgentLens answers the question other tools don't:

Why did my agent behave this way?

Not a prompt tracker. Not a token dashboard. A diagnostics platform for agent behavior — tool usage, memory access, workflow paths, decisions, failures, retries, latency, and cost.

pip install pyagentlens               # core: store + dashboard + manual API + auto-patch
pip install "pyagentlens[langchain]"  # framework adapters as extras (langgraph, crewai, …)
pip install "pyagentlens[all]"        # every adapter

Package name on PyPI is pyagentlens; the import name is agentlens.

from agentlens import monitor
monitor.start()          # auto-patches installed LLM SDKs — that's it

Full docs: see Guide.md — in-depth reference for humans and AI coding agents (when-to-use-what tables, every API signature, per-framework snippets, troubleshooting, and an agent cheat sheet).

agentlens ui             # dashboard at http://127.0.0.1:7180

No server to deploy. No cloud. No external services. Data lives in a hidden ./.agentlens/agentlens.db (SQLite, WAL).

Three ways to instrument (mix freely)

1. Zero-config auto-patch

monitor.start() detects and patches whatever is installed — OpenAI, Anthropic, Google Gemini, Groq, LiteLLM — so every LLM call becomes an llm span with tokens and cost, with no code changes.

2. Manual (framework-agnostic, always available)

import agentlens
agentlens.init(project="research-bot")

with agentlens.trace_session("chat"):
    with agentlens.trace_agent("planner", role="lead"):
        agentlens.log_decision(options=["search", "answer"], chosen="search", reason="needs fresh info")
        agentlens.log_memory("read", "user_prefs", hit=True)
        agentlens.log_tool("web_search", args={"q": "ai news"}, result=[...], retry_count=1)
        agentlens.log_llm(model="gpt-4o", input_tokens=1200, output_tokens=400)

A tool/LLM logged with an error= (or a trace_* block that raises) is automatically recorded as a failure.

3. Framework adapters

Plus a universal agentlens.adapters.wrap_tool(fn) / @traced_tool() that works with any framework or none.

The dashboard

• Overview — runs, success/failure rate, cost, p50/p95 latency.
• Runs → Run detail — the workflow execution graph (span tree: session → workflow → agent → tool/memory/llm/decision), click any node to inspect.
• Agent Explorer — runs, failures, latency, and attributed cost per agent.
• Tool Explorer — most used / slowest / most failed tools, retries.
• Memory Explorer — reads/writes/updates/deletes and read hit-rate.
• Failure Explorer — exceptions, messages, retries, jump to the failing run.
• Cost Explorer — token usage, cost by day, by model, most expensive agents.

Try it now (no keys needed)

pip install -e .
python examples/demo_agent.py
agentlens ui

Design

Local-first, zero-config, framework-agnostic, async-safe. Observability code never crashes your app (emit failures are swallowed). One Span model covers every event; kind-specific data lives in attributes. Cost is computed from an editable price book (agentlens/server/pricing.py).

CLI: agentlens ui [--port 7180] [--host] [--backend-store-uri PATH], agentlens providers, agentlens version. Override the port with AGENTLENS_PORT.

Single process writes straight to local SQLite. For multiple processes/workers sharing one dashboard, run agentlens ui once and point each worker at it: agentlens.init(project=..., tracking_uri="http://127.0.0.1:7180").

See Guide.md for the complete reference.