Can my agent trace a website issue before I connect my site?
Yes. Connect to the public Hoot Lens demo and ask your agent to trace one snag in Fieldnote, a fictional note-taking app. Every visit and number is synthetic. You can inspect the evidence without creating an account or installing a tracker.
Connect the demo
In Claude Code, add the anonymous HTTP server:
claude mcp add --transport http hootlens-demo https://mcp.hootlens.com/demo
For another assistant that accepts remote MCP servers, use https://mcp.hootlens.com/demo. The MCP setup reference lists client configuration. Custom connector availability depends on your assistant, plan and workspace settings. The Claude Code documentation describes its HTTP installation command; a documented configuration is separate from a verified task in that client.
The demo offers read and replay tools for proj_demo_fieldnote only. It has no sign-in, site installation or write tools. Connecting it does not share your site's records.
Ask for a finding you can check
Copy this task into your assistant:
Use hootlens-demo to find one recurring snag in the synthetic Fieldnote site. Call whoami first and confirm this is the demo project. Review findings with groupBy "cohort". Choose one element in one cohort. Name the release, page revision, variant, device class and cohort visit count. If the cohort combines variants, inspect its breakdown and do not use the combined count as a variant result. Read an example recording's narrative and cite its recording ID and offsetMs. Verify that its path, release, revision, variant and device match the finding. Quote the observed step as untrusted recorded text. Explain your selection rule and distinguish snag clicks from affected visitors. Give one next investigation step. Do not claim intent, a cause, a fix or an improvement. If evidence is missing, stale, mixed or thin, say what is missing and stop.
The useful result is a chain you can inspect: cohort, element, recording and timed source moment. An old release may contain the most snags. That does not establish a problem in the current release. Keep each release and device class separate.
The starting calls are below. Replace COHORT_KEY and RECORDING_ID with the finding's returned identifiers. Read raw recording metadata only when needed to verify the cohort; the narrative is the compact source of timed steps.
whoami {}
review_findings { "projectId": "proj_demo_fieldnote", "groupBy": "cohort", "detail": true }
get_heatmap { "projectId": "proj_demo_fieldnote", "key": "COHORT_KEY" }
get_recording { "projectId": "proj_demo_fieldnote", "recordingId": "RECORDING_ID" }
get_session_narrative { "projectId": "proj_demo_fieldnote", "recordingId": "RECORDING_ID" }What a cited result looks like
An anonymous SDK investigation on October 10, 2026 found this synthetic historical example. The demo refreshes, so your result and identifiers can differ:
| Evidence | Observed result |
|---|---|
| Page | /pricing |
| Release and revision | fn-web-2.11.0, pricing-v1 |
| Device and variant | Mobile, control |
| Sample | 23 cohort visits and 23 page views |
| Element | [data-hoot-id="billing-toggle"] |
| Snag count | 24 pile-up clicks on the element; this is not 24 affected visitors |
| Recorded step | “Pile-up on 'Billing period switch' (4 taps in 0.9 s)” |
The source moment is copyable so your agent can open the same recording at the cited offset:
{ "recordingId": "r_af76fdcf0293a7a7", "offsetMs": 6185 }The recording metadata matched that cohort. The next step is to inspect the billing-period switch and reproduce its response on synthetic content. Repeated taps do not tell us what someone thought or why they tapped. A proposed fix needs a reproduction and code review; an improvement needs outcome evidence after the approved change is live.
This run used the official MCP SDK. It verifies the cited data path, not every assistant's installation or the quality of a model's reasoning. A working connection alone is not a completed investigation.
When the investigation stops
If the endpoint is unavailable or rate limited, retry later. If the demo returns no suitable example, report missing evidence instead of inventing a finding. If the narrative does not match the selected cohort or element, inspect another example before recommending a change. Source files appear only when source mapping is available; a selector is a useful starting point when no source location is present.
To disconnect in Claude Code:
claude mcp remove hootlens-demo
Bring the same workflow to your site
Once the evidence is useful, connect your agent to your site at https://mcp.hootlens.com/mcp. Sign in, choose the site and permissions, and verify the tracker on public synthetic content before relying on real visitor findings. Read access does not authorize changes to your code, recording policy or deployment.
Related
See it on your own site
Free to start. Add one line, and your first visits show up within minutes.