AI Agent Verification · Browser Intelligence + MCP

Give Cursor a real browser verification gate—not a guess.

SiteLens combines browser automation, observability, diagnostics, visual regression, and evidence into one coherent workflow. AI Agent Verification is how coding agents prove a change works: SiteLens MCP runs Playwright browsers, executes flows from .sitelens/flows, then inspects console/network/diagnostics, compares runs, navigates the timeline, and reads SiteLens Reports. Agents aren’t limited to running prewritten tests — they can help create automation from recorded browser behavior, run it, inspect evidence, and iterate.

Cursor prompt library Record with agents Install SiteLens MCP Browser Intelligence
Cursor chat → MCP tool → Chromium on localhost
Flow pass/fail + screenshots — evidence agents can cite

What AI Agent Verification is

Shipped

A verification loop for AI coding agents: define repeatable browser flows as JSON, invoke them through SiteLens MCP from Cursor, and require a green run before marking work complete. Unlike ad-hoc “open the browser” instructions, flows produce structured step results, named screenshots, and plain-language failures agents can act on.

Structured flows

JSON documents under .sitelens/flows/{flowId}.json with clicks, fills, assertions, and network checks—up to 200 steps per local run.

Evidence, not vibes

Each run writes a coherent evidence model locally: summary, report, screenshots, console/network events, diagnostics, and timeline—artifacts agents can cite, not vibes.

Completion gates

Prompt-library patterns (below) tell Cursor to re-run flows until pass—closing the gap between “code compiles” and “user journey works.”

Create automation from recorded browser behavior

Codegen

Agents can help author SiteLens flows — not only execute them. Through MCP they start a headed Playwright Codegen session, watch status, finalize or cancel, convert the recording into a SiteLens flow, then run it and inspect Browser Intelligence evidence.

Start sessionHuman interactsFinalizeConvertRunInspect

MCP tools

  • sitelens_flow_codegen_start
  • sitelens_flow_codegen_status
  • sitelens_flow_codegen_finalize
  • sitelens_flow_codegen_cancel
  • sitelens_flow_codegen_convert

Be precise about headed recording

An MCP client cannot magically perform the human browser interaction by itself. Codegen opens a headed browser; a person (or local operator) interacts with the site while the agent orchestrates the session and conversion. After conversion, the result is a normal SiteLens flow under .sitelens/flows with an explicit complete/partial contract.

Prefer Desktop Record Flow when you want a visual review UI, or CLI sitelens flow codegen for terminal workflows. Same portable artifact either way.

How SiteLens works with Cursor

MCP

Install @sitelensapi/mcp, add the server block to ~/.cursor/mcp.json, restart Cursor, then ask the agent to call SiteLens tools by name. Local tools (sitelens_setup, sitelens_run_local_qa, sitelens_run_local_flow, compare/list helpers) run on your machine with no API key. Cloud tools activate when you set SITELENS_API_KEY for hosted runs and persisted history.

Step-by-step: SiteLens MCP product page · Cursor global setup guide · dashboard MCP cockpit (copy-paste JSON).

MCP overview

Local-first

Core tools for agent verification—each maps to a concrete Cursor prompt.

Setup & ad-hoc QA

  • sitelens_setup — verify Chromium and artifact dirs
  • sitelens_run_local_qa — quick desktop/mobile pass on a URL
  • sitelens_qa_run — lower-level local runner when needed

Flows & regression

  • sitelens_run_local_flow — run .sitelens/flows/{flowId}.json
  • sitelens_list_local_runs / sitelens_compare_local_runs — history + diffs
  • sitelens_compare_screenshots — pixel-diff two PNGs on disk
  • Codegen lifecycle — author flows from recorded browser behavior

Optional Cloud/API: sitelens_run_qa, sitelens_run_flow, sitelens_list_runs when SITELENS_API_KEY is configured.

Local flow execution

Project flows live beside your code. Cursor invokes sitelens_run_local_flow with a flowId (filename without .json) and optional URL override—Chromium runs on your machine and writes artifacts under your local SiteLens artifact directory.

1
Add a flow

Create .sitelens/flows/my-feature-smoke.json with version, url, viewports, and steps (assertText, click, fill, expectEither, …).

2
Prompt Cursor 
Use the SiteLens MCP tool sitelens_run_local_flow with flowId my-feature-smoke and url http://127.0.0.1:3000/ — report pass/fail and screenshot paths.
3
Gate completion

Use the Verification Completion Gate prompt (below) so the agent re-runs until the flow passes or reports blockers with evidence.

Flow authoring guide: MCP flows & templates.

Flow examples

Real repo flows

These flows ship in the SiteLens monorepo—copy patterns into your project or run them locally when developing SiteLens itself.

page-intelligence-marketing-smoke

Marketing homepage CTA & workflow smoke — asserts hero copy, nav labels, flagship templates, and primary CTA navigation to Desktop.

  • 17 steps · desktop viewport
  • Exercises in-page anchors and primary link navigation
  • Default url: local marketing preview (:4173)
sitelens_run_local_flow
flowId: page-intelligence-marketing-smoke
url: http://127.0.0.1:4173/

tldledger-name-to-bulk

Multi-step app journey: Name Ideas tab → generate names → select all → Bulk Check tab → run bulk check with flexible outcome assertions.

  • Tab navigation, fill, network wait, expected HTTP status, expectEither
  • Auth profile: tldledger
  • Default url: local app dev server (:5173)
sitelens_run_local_flow
flowId: tldledger-name-to-bulk
url: http://127.0.0.1:5173/

proof-homepage-smoke

Minimal proof flow—confirms the marketing homepage loads and exposes “SiteLens Platform” copy. Ideal starter template for agent verification.

  • 2 steps · fast sanity check
  • Catalog template id: proof-homepage-smoke
sitelens_run_local_flow
flowId: proof-homepage-smoke
url: http://127.0.0.1:4173/

Cursor Prompt Library

Public

Paste into Cursor chat as-is—each prompt names SiteLens MCP tools explicitly so agents do not substitute raw Playwright or terminal scripts.

Teach Cursor SiteLens

You are working in a project with SiteLens MCP configured. Before browser QA, call the SiteLens MCP tool sitelens_setup once per session. For quick URL checks use sitelens_run_local_qa; for repo flows use sitelens_run_local_flow with flowId from .sitelens/flows/. Do not run raw Playwright or curl unless I ask. Always summarize pass/fail, failed steps, and screenshot paths from the tool response.

Generate a SiteLens Flow

Read existing .sitelens/flows/*.json in this repo. Generate a new SiteLens flow JSON (version 1) that verifies [describe user journey]. Use assertText, assertSelector, click, fill, waitForText, and when verifying API failure paths use request/expectNetwork with expect.status (e.g. 403 or 500). Target url http://127.0.0.1:3000/ and keep under 40 steps. Save to .sitelens/flows/[flow-id].json and tell me the flowId to pass to sitelens_run_local_flow.

Verification Completion Gate

Before you mark this task complete, run the SiteLens MCP tool sitelens_run_local_flow with flowId [flow-id] and url [local-dev-url]. If any step fails, fix the issue and re-run until pass or report blockers with screenshot paths. Do not claim verification is done until the flow passes.

Coverage Gap Discovery

List .sitelens/flows/*.json and compare against recent feature changes in [area]. Propose 2–3 new flowIds or step additions to close coverage gaps. For each gap, draft a sitelens_run_local_flow prompt I can paste into Cursor and note which Page Intelligence fields (primaryActions, workflowStages) would help plan the flow.

Related

Contact & onboarding

Agent verification rollout, MCP help, or flow authoring — same short form everywhere.