AgentLens

See what your AI agents are actually doing.

Local-first tracing and debugging for AI agents — every LLM call, tool call, retrieval step and handoff, on a timeline you can actually read.

Light mode, and what a failure looks like

Why AgentLens?

An agent run is a tree: it plans, calls three tools (two of them concurrently), hands off to a sub-agent, retries something that failed, and answers. When the answer is wrong — or takes nine seconds — print() and a wall of logs will not tell you which step went wrong.

AgentLens records that tree. Two lines of code, one local command, and you can see every step, how long it took, what went in and what came out.

• Local-first. Traces go to a SQLite file on your machine. Nothing is uploaded.
• Framework-agnostic. The core knows nothing about any agent framework.
• Tiny API. A decorator and a context manager. That's the whole thing.
• Sync and async. Correct nesting across await points and concurrent tasks.

Quickstart

pip install agent-lense

The package installs as agent-lense and imports as agentlens — the shorter name was already taken on PyPI by an unrelated project.

git clone https://github.com/uoparaji/agentlens
cd agentlens
pip install -e .

Instrument a function:

from agentlens import trace, span


@trace
def research_agent(topic: str) -> str:
    with span("search", span_type="tool", metadata={"query": topic}) as s:
        results = search(topic)
        s.set_output(results)

    with span("summarize", span_type="llm") as s:
        s.set_model("claude-sonnet-4")
        answer = summarize(results)
        s.set_tokens(prompt_tokens=1_200, completion_tokens=180)
        s.set_output(answer)

    return answer


research_agent("quantum error correction")

Then look at it:

agentlens ui        # opens http://localhost:4180

No agent of your own yet? Generate a real trace in one command:

agentlens demo && agentlens ui

What you can inspect

Example

One run, as AgentLens records it:

research-agent                          1.58 s   agent
├── plan                                185 ms   llm         90 tokens
├── gather_sources                      423 ms   workflow
│   ├── web_search                      422 ms   tool
│   │   ├── GET arxiv.org               420 ms   custom     ← the slow one
│   │   ├── GET nature.com               91 ms   custom
│   │   └── GET acm.org                 131 ms   custom
│   └── vector_search                   110 ms   retrieval
├── analyst-agent                       329 ms   agent      ← handoff
│   ├── rank_sources                     41 ms   custom
│   └── synthesize                      287 ms   llm       110 tokens
├── fact_check                           83 ms   tool      ✗ ConnectionError
├── fact_check_retry                    121 ms   tool
└── final_answer                        432 ms   llm       140 tokens

The three GET spans ran concurrently — in the dashboard their bars overlap, and one of them is visibly three times longer than the others.

The API

There are three things to learn.

from agentlens import trace, span, current_span

@trace marks one agent run. It works bare, with arguments, on sync functions, on coroutines, and as a context manager.

@trace  # name comes from the function
def agent(topic): ...


@trace("research-agent", metadata={"v": 2})
async def agent(topic): ...


with trace("research-agent") as run:
    run.set_metadata(user_id=42)

A @trace-decorated function called inside another trace records a nested agent handoff rather than starting a second run.

span(...) marks one step inside a run. Spans nest automatically.

with span("search", span_type="tool", metadata={"query": q}) as s:
    s.set_input(q)
    s.set_output(results)
    s.set_model("claude-sonnet-4")
    s.set_tokens(prompt_tokens=1200, completion_tokens=180)

Built-in span types are agent, llm, tool, retrieval, workflow and custom — but any string works, and the dashboard colours the known ones.

current_span() reaches the active span from anywhere, which is handy deep inside a callback or a provider wrapper:

if s := agentlens.current_span():
    s.set_tokens(prompt_tokens=response.usage.input_tokens)

Exceptions are recorded automatically and re-raised untouched — AgentLens never changes your control flow.

Async

Tracing works identically with async def, and nesting stays correct under concurrency because context is tracked with contextvars, not globals:

@trace
async def agent(question: str) -> str:
    async with span("gather", span_type="workflow"):
        pages = await asyncio.gather(fetch(a), fetch(b), fetch(c))
    return await summarize(pages)


# Two runs at once produce two clean, separate traces.
await asyncio.gather(agent("q1"), agent("q2"))

Integrations

LlamaIndex

from agentlens.integrations.llamaindex import instrument

instrument()

index.as_query_engine().query("what changed?")

Captured automatically: every LlamaIndex span (query engines, retrievers, LLMs, embeddings, agents, workflows) with its own parent/child structure preserved, a span type inferred from the class and method, call arguments and return values, the model name and token counts reported by LLM events, and exceptions. LlamaIndex work started inside one of your own traces nests underneath it; otherwise it becomes a run of its own.

An error inside the integration is logged once and never propagates into your application.

Local-first and private

Agent traces contain prompts, documents and user data. AgentLens treats them accordingly.

• Traces are written to ~/.agentlens/traces.db (override with $AGENTLENS_DB).
• Nothing is sent anywhere. There is no account, no server, no telemetry.
• The dashboard binds to 127.0.0.1 by default.

Turn payload capture off, or filter it:

import agentlens

agentlens.configure(
    capture_io=False,  # keep timings, drop payloads
    redactor=agentlens.redact_keys(["api_key", "email"]),
)

A redactor is any (field, value, span) -> value callable, applied to every input, output and metadata value before anything is written to disk:

def redactor(field, value, span):
    return "<redacted>" if span.span_type == "llm" and field == "input" else value

Environment variables work too, which is useful in CI and shared machines:

AGENTLENS_DISABLED=1      # tracing becomes a no-op
AGENTLENS_CAPTURE_IO=0    # names, timings and tokens only
AGENTLENS_DB=./traces.db  # where to store traces

CLI

agentlens ui                 # launch the dashboard (--port, --host, --db, --no-open)
agentlens demo               # generate example traces  (--ui to open straight away)
agentlens list               # recent runs, in the terminal
agentlens clear              # delete every stored run
agentlens --version

How it works

your code ──▶ SDK (contextvars) ──▶ Tracer ──▶ Storage ──▶ SQLite file
                                                  ▲
 integrations ────────────────────────────────────┘
                                                  │
                                    FastAPI (read-only) ──▶ React dashboard

• SDK (trace, span) — tracks the active trace/span with contextvars.
• Tracer — builds Span and Trace models, times them, records errors. Storage failures are caught and logged; they never reach your agent.
• Storage — an abstract interface with a SQLite implementation. The tracing API does not depend on SQLite, so another backend can be added later.
• Server — a small read-only FastAPI app that also serves the dashboard.
• Integrations — adapters that use only the public SDK.

Roadmap

• CrewAI, LangChain/LangGraph and OpenAI Agents SDK integrations
• Cost estimation from token counts and model pricing
• Trace comparison (diff two runs of the same agent)
• Search across span inputs and outputs
• Export to OpenTelemetry

Contributing

Contributions are very welcome — especially integrations, since each one is self-contained. See CONTRIBUTING.md; the setup is:

pip install -e ".[dev]"
pytest

License

MIT — see LICENSE.