Sundae

Design reviews shouldn't disappear into chat.

Sundae puts the live page and ChatGPT beside an evidence board you control. Each finding points back to what the model saw. You choose what to accept and preview. Sundae runs a fresh capture before it marks a measured issue fixed.

Open Sundae to prepare a workspace, or go straight to the guaranteed published /demo workspace. The demo needs no login, connector, hosted auditor model, capture-provider key, or plugin.

Two ChatGPT routes have been verified with GPT-5.6 Sol and GPT-5.6 Terra: the desktop app's built-in browser, and ChatGPT Work Cloud at chatgpt.com. OpenAI currently documents only the desktop app. GPT-5.6 Luna has WebMCP disabled and will not discover Site Tools.

Challenge work added during the submission period

Sundae did not exist before this challenge. The submission period opened on August 25, 2026; this repository was created on August 29, 2026, and every commit in its history falls inside the submission window. The whole repository is challenge work.

Related MCP server: @metricspot/mcp-server

Why WebMCP

Sundae's page-hosted tools operate the evidence board the person can see and control:

1. ChatGPT measures the approved scope and reads the visible board.

2. It keeps measured facts, supported design judgment, and what was not seen distinct.

3. The person decides whether to accept, defer, or dismiss a finding and approves any preview.

4. Sundae renders a reversible preview and requires fresh matching evidence before a measured issue is called fixed.

5. Tool-named receipts leave each agent action inspectable on the same page.

Without WebMCP, this workflow collapses into a screenshot conversation, a hidden API report, or brittle click automation. WebMCP gives the agent specific commands while the person keeps control of the page.

Judge path

Site Tools have been verified on GPT-5.6 Sol and GPT-5.6 Terra in two places: the ChatGPT desktop app, and ChatGPT Work Cloud at chatgpt.com. OpenAI's Site tools documentation currently names only the desktop app. GPT-5.6 Luna has WebMCP disabled and will not discover Site Tools. Site Tools are also unavailable in Enterprise and Edu workspaces.

1. Open ChatGPT Desktop's built-in browser, or ChatGPT Work Cloud's browser, at https://usesundae.vercel.app/demo.

2. Click Site tools in the browser address bar. You should see 11 Sundae tools. If the panel is empty, check the model first.

3. Ask: "Audit this page with its Site Tools. Keep measurements, judgment, and what you did not see separate, and ask me before any decision or preview."

4. Watch the board: measured evidence and receipts appear as tools run, and the Agent tool calls counter in the top bar increments.

5. ChatGPT asks before changing a decision or starting a preview.

6. Accept a finding with a visible reason, then let it run preview_fix and verify_recapture.

7. Confirm the board shows the fresh measurement, the verification state, and tool-named receipts.

A host may deny an individual Site Tool call. In a production run, ChatGPT's auto-review blocked record_audit_brief with "Browser Use rejected this action due to browser security policy". Sundae never saw that call. The rest of the audit still completed.

In an ordinary browser or with an unsupported model, Sundae still provides every deterministic measurement and every human control. The top bar reads Human controls ready, and the Agent tool calls counter stays at 0.

Evidence contract

Sundae separates:

• Measured: deterministic observations from the rendered page, accessibility tree, or inspected WebMCP contract.

• Judged: evidence-linked UI, UX, or Interaction critique for the visible product job.

• Not seen: routes, states, or motion windows outside the captured evidence.

• Decision: a reversible state chosen by the person.

• Preview: a temporary visual proposal. It does not prove that the product changed.

• Verification: a fresh same-scope measurement. A judged finding remains unverified unless it is reassessed.

ChatGPT performs the design critique through Site Tools using visible screenshot and page evidence. Sundae stores the category and product-job tag; it does not auto-classify an industry or call a server-side model.

One group is one fix

A product page can yield hundreds of measurements. A list of 91 findings is only a census. Sundae groups measured findings by the change that would resolve them, then ranks each group using severity, rendered area, and distance from the fold.

• Contrast groups by the exact foreground/background colour pair. One pair is one design-token fix however often it recurs.

• Tap targets group by shape class because an icon control, inline link, and button need different CSS fixes.

• Everything else stays per-instance.

Each group reports its instance count and shows the worst instance: 44 instances · worst shown. Controls that are invisible, offscreen, or smaller than 8 × 8 CSS px are dropped before grouping. The same grouping rules apply to public pages and the demo fixture.

Contrast is measured against a composited background. Sundae walks the ancestor chain and blends every translucent layer rather than stopping at the first non-transparent colour. Without this, the engine reads a 4% white overlay as pure white and can flag legible text. On one production dark-theme site, that error reported 17.51:1 text as 1.06:1.

On /demo, preview_fix renders a pre-authored improved variant of the controlled fixture. The receipt attributes that fix to the fixture, not the agent. Then verify_recapture measures the improved state and reports which accessible-name, tap-target, contrast, and horizontal-overflow findings remain. On a public checkpoint, the agent supplies bounded CSS.

Page-hosted tool surface

The five tools a judge will see

audit_current_scope → get_board_context → record_visual_finding → set_finding_decision → verify_recapture

Supporting tools

The included /demo registers eleven Sundae workbench tools:

ChatGPT's built-in browser does not discover tools registered inside iframes. Sundae's 11 tools are registered at the top level, so the count shown in Site tools is exact. The audited fixture's own tools live inside the iframe. inspect_agent_surface reads those contracts because the host cannot reach them directly.

A public workspace adds four bounded capture commands, for 15 tools total:

Each input is a closed, bounded schema. Long-running capture work receives the invocation AbortSignal. Mutating commands are not marked read-only; page- or capture-derived output is marked untrusted. Registration uses one abortable transaction, and human controls call the same command implementation as Site Tools.

Public capture boundary

Sundae audits public HTTPS pages its configured browser provider can render. When a page blocks automated rendering, requires login, or exceeds browser limits, Sundae records what was not seen instead of claiming a complete audit.

Typing a public URL and pressing Enter approves that exact target for the current browser session and starts capture. ChatGPT may then call capture_public_page on the same URL. Additional same-origin routes still need a human-named journey step.

Public capture opens the approved page in a Cloudflare browser session and runs Sundae's own audit engine inside that page with page.evaluate(). Measurement happens in one browser session. A small page settles in under ten seconds; a large one, where the design-signal walk reaches its 1,500-node cap, takes closer to thirty. Sundae never accepts credentials in URLs, private-network targets, nonstandard ports, target-site cookies, silent tabs, login flows, form submission, recursive crawling, or off-origin navigation. Existing board evidence remains intact when the provider fails.

For local public-capture development, configure opaque values outside source control:

SUNDAE_BROWSER_WORKER_URL
SUNDAE_BROWSER_WORKER_SECRET
CLOUDFLARE_ACCOUNT_ID
CLOUDFLARE_API_TOKEN
CAPTURE_GATE_SECRET
SUNDAE_APP_ORIGIN

Development

npm install
npm run dev
npm run check
npm run build

npm run check runs TypeScript, Oxlint, the product token gate, format checking, the full test suite, and the production build. The interface uses the repository token contract and keeps the cyclomatic-complexity ceiling ratcheted.

Key paths:

• app/: landing page, /demo workspace redirect, controlled fixture, and capture API.

• components/Workbench.tsx: shared command implementation and audit state.

• lib/webmcp/: page-hosted tool contracts, registration, and bounded results.

• lib/audit/: measured findings, structured judgment, contract inspection, and recapture comparison.

• lib/capture/: URL policy, Cloudflare snapshot adapter, evidence extraction, and failure boundaries.

• tests/ and evals/: deterministic product, contract, and prompt regressions.

License

AGPL-3.0-only. See LICENSE.