Trace the browser-agent stack you already use

BrowserTrace records the screenshots, URLs, actions, model inputs, model outputs, statuses, and errors around failed browser-agent runs.

View repo Open live demo

Improving an adapter note or docs example? Use the First PR Recipe to keep the first contribution small and reviewable.

Browser Use Stagehand Skyvern Playwright + LLM Computer Use

BrowserTrace timeline preview for AI browser-agent debugging

AOS mapping research

This is research-only mapping, not an AOS compliance claim. BrowserTrace is tracking how browser-agent artifacts could map to OWASP Agent Observability Standard concepts while upstream browser/GUI artifact guidance is still open.

Tracker: BrowserTrace issue #237.

BrowserTrace field	Closest AOS shape	Status	Current note
Run id and step id	Step context plus tool execution or correlation id	partially mapped	BrowserTrace has run ids and ordered steps; an export-level AOS correlation convention is still needed.
Action label or tool call	`steps/toolCallRequest`	partially mapped	Action labels and integration method metadata exist, but normalized tool ids and execution ids are not standardized yet.
Status and error	`steps/toolCallResult` with `isError`	partially mapped	Document the research-only `toolCallResult` result shape as run status, step status, error message/type when available, action/tool label, step id, and public export/redaction state; this remains a research mapping, not an AOS compliance claim.
Screenshot or future video artifact	`FileWithUri` preferred over inline `FileWithBytes`	research gap	BrowserTrace keeps screenshots as local artifacts; browser screenshot/video examples are still unresolved upstream.
Current browser URL	Separate event or resource attribute	research gap	Keep URL separate from screenshot evidence so future mappings can redact each field independently; the AOS attribute placement is still open.
Model input and output	Message, reasoning, or linked model-message summary	research gap	Use compact summaries where possible, but do not force full prompt or model I/O into every tool-result payload.
Public export redaction state	Artifact and field-level metadata	research gap	Future mappings should preserve explicit omitted or redacted markers instead of silently dropping sensitive evidence.

Browser Use

Attach BrowserTrace to a Browser Use agent through its step callback surface.

Open the Browser Use debugging guide for a focused quickstart, export commands, and public-safe trace workflow.

from browser_use import Agent
from browsertrace import Tracer
from browsertrace.integrations.browser_use import attach_tracer

tracer = Tracer()
agent = Agent(task="...", llm=ChatOpenAI(model="gpt-4o"))

with attach_tracer(agent, tracer, name="browser-use run"):
    await agent.run()

Stagehand

Wrap a Stagehand page so high-level browser-agent calls become trace steps.

Open the Stagehand debugging guide for the wrapper quickstart, public-safe export commands, and adapter feedback link.

from stagehand import Stagehand
from browsertrace import Tracer
from browsertrace.integrations.stagehand import wrap_stagehand

tracer = Tracer()
stagehand = await Stagehand(...).init()
page = wrap_stagehand(stagehand.page, tracer, name="stagehand run")

await page.goto("https://example.com")
await page.act("click the login button")
await page.extract("get the headline")
page.bt_run.close()

Skyvern

Wrap a Skyvern-shaped client to record task and workflow calls without adding a hard Skyvern dependency.

Open the Skyvern debugging guide for task/workflow tracing, public-safe export commands, and adapter feedback link.

from skyvern import Skyvern
from browsertrace import Tracer
from browsertrace.integrations.skyvern import wrap_skyvern

tracer = Tracer()
skyvern = wrap_skyvern(Skyvern(...), tracer, name="skyvern run")

await skyvern.run_task(
    url="https://example.com",
    prompt="extract the invoice total",
    wait_for_completion=True,
)

skyvern.close()

Playwright + LLM

Use snapshots around the points where your LLM chooses the next browser action.

Open the Playwright + LLM debugging guide for a focused trace workflow and public-safe export commands.

from browsertrace import Tracer

tracer = Tracer()

async with tracer.run("playwright agent") as run:
    await page.goto("https://example.com")
    await run.snapshot(page, action="opened example.com")

    decision = await llm_choose_next_action(...)
    await page.click(decision["selector"])
    await run.snapshot(
        page,
        action="clicked selected element",
        model_input=decision["prompt"],
        model_output=decision,
    )

Custom computer-use agents

Place snapshots at the observe, decide, and act points in your own browser-agent loop.

Open the computer-use debugging guide for a generic wrapper pattern, export commands, and public-safe trace workflow.

from browsertrace import Tracer

tracer = Tracer()

async with tracer.run("custom computer-use agent") as run:
    await page.goto("https://example.com")
    await run.snapshot(page, action="opened task page")

    observation = await describe_page(page)
    decision = await model.choose_next_action(observation)

    await page.click(decision["selector"])
    await run.snapshot(
        page,
        action=f"clicked {decision['selector']}",
        model_input=observation,
        model_output=decision,
    )