Let AI agents use your website.Prove every tool before release.

Turn real actions into WebMCP tools. Test in Chrome. Track tool health without payloads.

siteverbsiteverb / stateful fixtureRP
Sample fixture · runtime signalTool analytics
Last 7 days⌄
Tool calls4,161
Success rate96.9%
Failed calls96
P95 latency494 ms
Calls over timeSample callback metadata
Top toolsSee all ›

catalog.search-products1,814

cart.add-item1,067

cart.get-current793

Recent releasesSee all ›

PR #8 · failed1 journey

PR #7 · passed1 journey

main · currentv0.1.0

WebMCP testing and analytics.Evidence agents and teams can trust.

Understand

What is WebMCP?

WebMCP is a proposed web standard for exposing structured website actions as tools that AI agents can discover and invoke. It replaces brittle click sequences with explicit names, schemas and results.

Read the Chrome overview →

Verify

How do you test WebMCP tools?

Siteverb audits the declared contract, discovers tools in a named browser and runs deterministic journeys with explicit postconditions before a release can pass.

See the conversion workflow →

Measure

What is WebMCP analytics?

Siteverb groups calls, outcomes and bounded latency by stable tool ID. It does not collect prompts, arguments, results, DOM, credentials or user identity.

Review the data boundary →

Create. Prove. Improve.Your code stays yours.

01

Create agent-ready tools

$ npx skills add siteverb/siteverb --skill siteverb

✓ WebMCP surface mapped

✓ 6 stable tools declared

✓ 1 journey ready to verify

The Siteverb skill maps real app actions, wires the SDK and writes one reviewable siteverb.webmcp.json contract.

02

Prove every release

Siteverb WebMCPPassed
Release evidence passed

1 passed · 0 failed · 0 skipped

Browser
Chrome 151 · real browser
Audit
0 errors · 0 warnings
Impact
3 observed calls

Your GitHub Action runs deterministic journeys in a named browser. Siteverb keeps one release Check current.

03

Improve from call health

Sample tool healthLast 7 days
Tool calls4,161
Success96.9%
P95494 ms

See calls, outcomes and latency by stable tool ID—without prompts, arguments, results or DOM.

Know what broke.Know which tools get used.

Health for every tool

Calls, outcomes and bounded latency, grouped by the stable ID already in your contract.

Payloads stay private

Prompts, arguments, results, DOM, identities and raw errors have no field in the event contract.

Metadata only
Prompts×Arguments×Results×DOM×User IDs×Raw errors×

Context on every failure

Failed journeys name the affected stable tool and recent call volume—never a causal claim.

Pull request #28Checkout journey failed

cart.prepare-checkoutaffected tool

487calls / 7d

One release decision

New commits replace stale evidence. Duplicate deliveries never create a second Check.

Siteverb WebMCP Passed

Open before hosted

The skill, SDK, contract, audit, runner and Action work without an account. The cloud adds history, not lock-in.

Skill→SDK→Contract→Action→Cloud

Live staging evidence

A real release loop.

Passing and failing pull requests, plus one idempotent SDK batch, meet in the same console.

3Observed tool calls
100%Fixture success rate
250 msP50 latency
1.0 sP95 latency

Deliberate acceptance failureFind a trail shoe and prepare checkout

Affected stable tool
catalog.search-products
Previous 7 days
3 calls · 3 completed
Payload content
Excluded

Trust boundary

Your app acts.Siteverb keeps the evidence.

Registration only exposes a tool. Your server still authorizes every action. Siteverb never becomes an agent or execution plane.

Read the security model
Customer website@siteverb/webmcpTools execute here
metadata only
Siteverb CloudExact aggregatesEvidence lives here
context only
Pull requestOne release CheckDecision appears here

Questions

Clear answers.Strict boundaries.

Does Siteverb run my application code?+

No. Your repository’s own GitHub Action runs deterministic browser journeys. Siteverb validates the bounded report and updates one Check.

What does the SDK send?+

Stable tool identity, schema hash, release, environment, outcome and duration. Prompts, arguments, results, DOM, credentials, full URLs, user IDs and raw errors are excluded from the event contract.

Can Siteverb tell which AI client called a tool?+

Siteverb records exactly which stable tool ran. It does not guess whether the caller was ChatGPT, Claude or a specific person because browser callback metadata is not trustworthy identity. Referral traffic and tool invocation are separate signals.

Which AI clients can use the tools?+

Chrome 152 has a Siteverb real-browser pass. ChatGPT Site Tools and Edge 150 use dated documented profiles. Additional clients are added only when evidence exists.

What happens when a pull request breaks a journey?+

The GitHub App fails the Siteverb Check and adds the previous seven days of observed calls for affected stable tools as context. It never claims the pull request caused those calls.

Can I use Create without the cloud?+

Yes. The open skill, SDK, contract, audit, runner and customer-owned Action remain usable without an account or hosted connection.

Is Siteverb ready for customer production data?+

Not yet. The private alpha is limited to allowlisted staging fixtures until live uninstall, deletion, retention and privacy-policy gates are complete.

Private alpha

See every tool clearly.

Review release evidence and call health in one place.

Checking stagingProduction collection remains disabled