UI Verify MCP Server

Triage one build, resolved by commitSha, prNumber, or buildId (exactly one). Returns the gate verdict and counts (total/changed/failed/unchanged), then the FIRST page (25) of the changed stories (story id, viewport, % pixels changed, review decision, the AI verdict aiVerdict=intended|regression + aiConfidence, and for a regression the judge's one-line aiFlagReason=what-looks-unintended, when AI review is on, and the story's own review-comment tally comments {total, unresolved}) and the first page of failed stories (failedStories[].kind is `render` when the story produced no frame, or `interaction` when it rendered and its play() assertion threw; failedStories[].error is the message's first line, errorTruncated marks the ones that carried more). When AI review ran, aiReview tallies how many changes look like regressions vs intended, and the changed list is ordered likely-regressions-first. changedNextCursor / failedNextCursor are non-null when there are more than 25: page the rest with list_build_stories. counts.unchanged is how many stories did NOT change and have a baseline - this triage call does NOT list them (keeps the response lean); call list_build_stories status=unchanged to browse them (e.g. to confirm a passed build's stories are 'identical to baseline'). comments is the build-wide review-comment tally {total, unresolved}: unresolved>0 means a designer/QA (or another agent) left an open note on a story - read it with list_comments BEFORE you accept. No image URLs here - call get_diff or render_diff_image for pixels.

Page through one build's stories filtered by status (resolved by commitSha/prNumber/buildId). Use it to read past get_build's first page, or to browse the unchanged/passed stories get_build only counts. status='changed' returns changed stories (same shape + order as get_build, regression-first, each with its comments {total, unresolved}; unresolved>0 means an open review note - read it with list_comments first); status='failed' returns failed stories; status='unchanged' returns the stories that did NOT change this build but have a baseline on this branch (storyId, viewport, browser) - fetch any of their images by storyId with get_diff or render_diff_image to confirm 'identical to baseline'. Returns { stories, nextCursor }: pass nextCursor back as `cursor` for the next page; null means no more. limit defaults to 25 (max 100). Unchanged is only available once the build has settled (an in-progress build has rendered nothing, so it returns an empty page).

Soft-delete a comment (get its id from list_comments). An agent can delete only agent-authored comments, never a human's. Deleting a thread root removes its replies too.

Replace the body of a comment (get its id from list_comments). An agent can edit only agent-authored comments, never a human's. Reference: https://uiverify.ai/docs/mcp-tools

Accept ALL changed stories in one build at once (resolved by commitSha/prNumber/buildId), advancing each story's per-branch baseline so the next build is clean. This is the bulk form of review_diff='accept' — use it when you've decided the whole build's changes are intended. It accepts every changed story, including stories that already have a prior denied review; failed stories are unaffected. Returns the build triage after accepting (first page of changed stories, each now showing decision='accepted'; page the rest with list_build_stories if changedNextCursor is set). Refused, with nothing accepted, while any changed story has an unresolved review comment: the error names those stories; read them with list_comments, address each note, resolve_comment its thread, then accept again.

Recent UI Verify builds for one project, newest first. Each row is a one-line verdict (gateStatus: passed | changed | failed) for a commit/PR — use it to find the build to inspect. No image URLs or per-story detail here: call get_build for the changed-story list, get_diff for images.

The cumulative 'this PR vs base' visual changeset (resolved by commitSha/prNumber/buildId) - what the whole PR does to the UI versus the branch it merges into, INDEPENDENT of what is left to review on the latest build. Unlike get_build (this commit vs the branch's own accepted baseline), this survives in-PR accepts: a story accepted mid-PR still shows under `changed` with status=accepted, so it answers 'what did this PR change' even after the gate reads clean. Returns base (the branch it merges into), counts {new,changed,removed,unchanged}, and the first page of `new` + `changed` + `removed` stories (storyId, title, name, kind, review status, aiVerdict). `removed` lists stories that HAD a baseline on the base branch but are gone from this PR's head (a deletion - otherwise invisible; each carries lastBuildId, the build that rendered its now-orphaned baseline). newNextCursor / changedNextCursor / removedNextCursor page the rest with list_pr_stories. No image URLs - fetch pixels by storyId with get_diff / render_diff_image.

Every review comment on a build's changed stories (resolved by commitSha/prNumber/buildId), each carrying its diffResultId, author (human name or 'agent'), body, createdAt, and anchor - where on the candidate image it points: null (whole diff), {type:'point',x,y}, {type:'rect',x,y,w,h}, or {type:'path',points,closed}, all normalized 0..1 to the candidate image. Use it to read what a designer/QA (or another agent) flagged on a screenshot before acting.

Per-story diff detail for a build (resolved by commitSha/prNumber/buildId). Returns diff metrics and presigned, time-limited URLs (download them to a file, or link them in a PR comment) for the baseline, candidate, and diff PNGs. Defaults to the changed stories; pass storyId for one specific story. When that storyId is an UNCHANGED story it returns its baseline (diffResultId null, changed false, the baseline URL as both baselineUrl and candidateUrl) - the story rendered identical to baseline. Use render_diff_image instead when you want the actual pixels inline for a vision model, not a URL. When AI review is on, each diff carries the judge's call: aiVerdict (intended|regression), aiConfidence, aiSummary (what changed), aiReasoning, and aiFlagReason; all null when AI review didn't run for it. Each diff also carries comments {total, unresolved} for THAT story - unresolved>0 means an open review note; read the thread with list_comments before accepting.

Leave a review comment on one of a build's changed-story diffs (resolve the build by commitSha/prNumber/buildId, target the diff by diffResultId from get_build/get_diff). The comment is authored as the agent. Optionally anchor it to a spot on the candidate image with `anchor`: omit or null for a whole-diff comment, {type:'point',x,y} to pin a point, {type:'rect',x,y,w,h} for a box, or {type:'path',points:[{x,y},...],closed} for a shape - all coordinates 0..1 normalized to the candidate image. `side` picks the pane ('after' the candidate, default; 'before' the baseline). Pass `parentId` (a root comment id from list_comments) to REPLY in that thread instead - a reply inherits the root's side and carries no anchor, so `anchor`/`side` are ignored. A new root comment is refused when a newer build exists on the branch (comment on the latest build); replies stay allowed. Reference: https://uiverify.ai/docs/mcp-tools

Every UI Verify project you can access, across all the teams you are a member of, most recently built first. Each row carries the projectId to pass to list_builds, or to a commitSha/prNumber lookup, when you can access several projects, plus the team and project slugs, name, type (storybook | playwright | vitest | bun | screenshot), repo, and lastBuildAt. Calls that take a buildId, diffResultId, or commentId never need a projectId. When you aren't in any project yet, it lists UI Verify's demo project instead (demo: true): every read tool works on it, so you can show what UI Verify does on a real build, but nothing can be accepted or commented there. Every response about the demo carries a note saying so. When the list is empty, note says how to get a first project.

Fetch a story's image as an INLINE image (base64 pixels, not a URL) so a vision model can look at it directly. Needs a native MCP client that renders image content; piped through raw curl it is useless (use get_diff's URLs there instead). Two ways to address it: (1) diffResultId + which for a CHANGED story - which is baseline|candidate|diff (the triptych) or before_after, the baseline and candidate SIDE BY SIDE (before on the left, after on the right) cropped to the changed region - and when a story changed in SEVERAL far-apart places, one such crop PER region stacked top to bottom, so a header-plus-footer change is two tight crops, not a page-tall image. before_after is usually what you want for a code change - it zooms to what moved instead of a full page. Get diffResultId from get_build (changedStories[].diffResultId) or get_diff. (2) a build selector (commitSha|prNumber|buildId) + storyId for ANY story's current image - the candidate if it changed this build, else its baseline. This second form is the only way to see an UNCHANGED/passed story's pixels (list them via list_build_stories status=unchanged), so you can confirm 'identical to baseline'. Errors if the image doesn't exist (a genuinely new story has no baseline, so no before_after either). When the story has unresolved review comments, a text note saying so comes with the image: read them with list_comments before you accept.

Mark a comment thread resolved or reopen it (get the id from list_comments). Resolving stamps the thread ROOT, so pass a root comment's id - a reply id is a no-op. Any team writer may toggle this.

Record a review on a single diff result (get diffResultId from get_build/get_diff). 'accept' makes the candidate the new baseline for that story on its branch, so the next build is clean; 'deny' records the rejection without changing the baseline; 'ignore' excludes the story from the gate persistently (the snapshot keeps differing on future builds but no longer flags the check) — use it for an intentional, ongoing diff like an animation or a live timestamp. This is the same accept/deny/ignore as the dashboard, attributed to the agent (and to you, when you logged in). Use after you've inspected the diff (e.g. with render_diff_image) and decided intended vs. regression. accept and ignore are refused while the story has an unresolved review comment (read it with list_comments, address it, resolve_comment the thread, then retry); deny is always allowed.

Page through the PR-vs-base changeset stories of one kind (resolved by commitSha/prNumber/buildId) - use it to read past get_pr_changeset's first page. kind='new' returns stories the PR adds that the base branch has no baseline for; kind='changed' returns stories whose image differs from the base branch's accepted image (even if accepted mid-PR, with status=accepted); kind='removed' returns stories that HAD a baseline on the base branch but are gone from head (a deletion; each carries lastBuildId, not a review status or verdict). Returns { stories, nextCursor }: pass nextCursor back as `cursor` for the next page; null means no more. limit defaults to 25 (max 100).