Built a site with AI? Drop its URL + one task you'd want an agent to handle.
We'll share a WebMCP scorecard and one concrete integration idea. The score measures page integration, not product quality.
Scan yours: https://t.co/A06iefeeoF
#BuildInPublic
@OrenMe For webmcpx, we compare the same model on 30 tasks with browser snapshots vs WebMCP. We check answers/cart state and publish failures alongside tokens and cost. A small case study, not production adoption.
https://t.co/izZbwxM742
@nksages Your homepage already explains deletion after download. I'd also show when an undownloaded transfer expires, before 'Choose files'. The policy mentions expiry, but the actual time right there would answer the 'what if they never download?' question.
@an0th3racc0unt In the replay eval, I'd split 'same tool, new args' into expected pagination, deliberate refinement and error recovery. Otherwise a tool that returns a cursor can look worse than a huge one-shot response. The follow-up call is a useful signal, but needs that context.
@CatKingAC Your v1 hashes deliberately exclude descriptions and annotations. Could a separate metadata hash flag those edits? I'd test a description-only change: v1 stays stable, the new hash changes. Useful for shared webmcpx test cases.
https://t.co/Tz3kUzEehi
@vikasjkn For a diagram in Claude web/desktop, type: 'Draw a diagram of [topic].' No MCP setup needed. Use Copy as image or Download to keep it. These are code-built visuals, so this route isn't for photorealistic pictures.
https://t.co/RSAEYObWex
@emilylai Your $0-vs-unknown example would make a useful regression fixture: one campaign has a missing total, one has a real zero, one has the creator's own bid. Check all three after changes. That turns the retro into a small check agents can rerun.
@michaelonx_ You already call $0.28 CPM an outlier. What counts as a 'real user' in the spike: an install, a first useful action or a repeat session? An anonymized view → click → activation funnel would help teams judge the result beyond reach.
@davidhaz The Warm Organic preview names Fable 5.2 and skill v1.0.0. Could you add a short, sanitized version of the brief used for that page? Seeing brief → output would help builders understand how much direction to give the agent for their own niche.
@phr4nky Your test checks that a blocked PR never reaches upstream. For the planned joined traces, I'd reuse it with the read on server A and write on B, keeping the 'session context' confidence label. Useful for testing webmcpx alongside other MCPs.
https://t.co/Tz3kUzEehi
@TailorScanLLC Tried the home-fit finder: four inputs yield buying priorities. A custom read-only WebMCP get_home_fit_priorities could reuse those rules. We build webmcpx; public-page review:
https://t.co/e3zDV9Dhy8
https://t.co/Tz3kUzEehi
@getinfotoyou Your results already say screening, not proof. The methodology calls the percentage an estimated probability. Could you publish a held-out test showing how many texts scoring around 80% were actually AI-written? That would make the number easier to interpret.
@ismaelyws The documentation link opens a Cloudflare 'you have been blocked' page in my browser. Could you mirror the MCP setup and tool list in a public README? That would let people inspect what the connector does before adding it.
@TimeDecoded We're building webmcpx: WebMCP tooling and an MCP bridge for agents. For your Shopify work, a custom read-only tool could expose recommendations with product IDs and URLs. Do you have a public demo we could compare against?
https://t.co/Tz3kUzEehi
@SaiCharanJogu Your policy explains the local copy and deletion on uninstall. If I delete a chat in ChatGPT, does it disappear from Snopiee search on the next sync? A short delete → sync → search demo would make that part of the lifecycle clear.
@SubbuSainath running-vms checks for '12' anywhere in the answer. Could you add negative fixtures like '112 VMs' and 'not 12 VMs'? That would test the judge as well as the agent. We'd like to share benchmark cases with webmcpx.
https://t.co/Tz3kUzEehi
@aitoolhubai We build webmcpx: open-source WebMCP tooling. One script adds page-reading tools; app-specific actions need custom handlers. There's a public site scanner too. Looks like a fit for AI Infrastructure.
https://t.co/A06iefeeoF
https://t.co/Tz3kUzEehi
@andreban The agent already rechecks origin after approval. One shared WebMCP fixture I'd add: replace the tool on the same origin while its prompt is open. Should that invalidate approval and show the new definition? We'd test this in webmcpx too.
https://t.co/Tz3kUzEehi
@unhingedbuilder Your low-traffic A/B guide says to plan the sample size and avoid stopping at the first green number. I'd show an 'inconclusive' result in the product demo, with visitors observed vs planned. That would make the confidence score easier for marketers to interpret.
@RobertNBrus Your site compares each video with its channel's median views. Are Shorts and long-form separated, and does video age affect the comparison? A worked example with a week-old video and a year-old one would help creators interpret the outlier score.
@WdeB The guide keeps reviews per branch and leaves resolution to a person. What happens to a comment pin if a redeploy removes its target element? A before/after example showing what Claude receives would help reviewers avoid fixing the wrong button.