MCP App Store
by KeenEthics
Brand API icon

Brand API MCP Server for ChatGPT

By Taste Labs

Overview

Brand API turns a website URL into a structured brand system: colors, typography, layout and components, plus the captured HTML, CSS and screenshots. Search a curated corpus of brand sites by aesthetic, build new pages inside an existing brand, and score how closely a page follows it. Ships with three skills: brand-search, taste-director and brand-adherence.

How to connect

  1. ChatGPT: Open the Brand API listing in the ChatGPT app directory and choose Connect. ChatGPT listing Brand API exposes 18 tools here.

Tools

extract_brand

ChatGPT
Extract the brand system of a website from its URL. Use the result to replicate a site, build new on-brand pages or whole websites that follow the brand, or find visually similar brands. Starts an asynchronous extraction job and returns immediately. Typical flow: 1. extract_brand(url) -> returns a brand_submission_id. 2. poll_brand_extraction(brand_submission_id) -> poll until status is "completed" (or "failed"). 3. get_brand_extraction_result(brand_submission_id) -> read the brand system. Pass sections=["colors", "typography"] to extract only those sections. Pass sections=["brand_assets"] for hosted, judged campaign images and semantic labels from external sources. Enabled full extractions include this section. Dependencies still run when needed; layout-dependent selections may take similar time to a full extraction. Omit sections to extract the full brand, which is required for search_similar_brands. Caching: a recent full extraction of the same URL can be reused, including for a selective request (only requested sections are returned). Selective results are stored for retrieval but never used as cache sources. poll_brand_extraction reports cache_hit: true on reuse. A cache hit is faster but NOT free -- every extract_brand consumes 2 credits, cache hit or not. Pass force=true to force a fresh re-extraction (also billed). To re-read a result you have already paid for without spending again, call get_brand_extraction_result with its existing brand_submission_id (or find it via list_brand_extractions) instead of submitting the URL again. Local pages: a server on localhost is not reachable from the Taste Engine. Expose it over the Tastelabs SSH tunnel (steps in the server instructions) and pass the printed http://crawl-<uuid> URL. Tunnel aliases are never cached, so each call re-extracts whatever is published behind the alias. Returns: {submission_id, status, source_url, accepted_at}. submission_id is the brand_submission_id to pass to the other tools. status begins as "accepted" and advances through queued/crawling/extracting.

extract_brand_asset

ChatGPT
Extract the design system of one brand asset from the URL of its image. A brand asset is one finished marketing piece: a billboard, a poster, an Instagram story or post, a social link card, or a display ad. The image can be the flat creative, a mockup, or a photo of the piece where it runs. The result is measured from its pixels: palette and gradients, type with an open stand-in font, layout, logo, elements, and people. Use it to make new pieces that follow how the brand designs its campaigns. Starts an asynchronous job and returns immediately. Typical flow: 1. extract_brand_asset(url) -> returns a brand_submission_id. 2. poll_brand_extraction(brand_submission_id) -> poll until status is "completed" (or "failed"). 3. get_brand_asset_result(brand_submission_id) -> read the design system. Costs 2 credits, refunded when the URL is refused or the extraction fails. Every call is a fresh extraction. A failed submission's error starts with NOT_A_BRAND_ASSET or UNREADABLE_IMAGE (send another image) or EXTRACTION_FAILED (retry). Some hosts refuse automated downloads ("URL answered 403"), and a Tastelabs tunnel alias (http://crawl-...) does not work here: host the image at a public https URL. Returns: {submission_id, status, source_url, accepted_at}. submission_id is the brand_submission_id to pass to the other tools.

extract_brand_kit

ChatGPT
Extract a brand kit from the URL of a brand guidelines PDF. The kit is the brand as its own guidelines state it: profile and voice, the palette with its proportions, typography, logo variants and misuse, photography, graphics, layout, buttons, cards, motion, and the book's rules. Use it when the source of truth is the guidelines document, not a website, to make new assets and pages that follow it. Starts an asynchronous job and returns immediately. Typical flow: 1. extract_brand_kit(url) -> returns a brand_submission_id. 2. poll_brand_extraction(brand_submission_id) -> poll until status is "completed" (or "failed"). 3. get_brand_kit_result(brand_submission_id, sections=[...]) -> read the kit. Costs 2 credits, refunded when the extraction fails. Every call is a fresh extraction. The URL must give the PDF file itself: a share page (Google Drive, Dropbox, Notion) gives a web page, and the submission fails. A failed submission's error starts with UNUSABLE_URL (the URL is blocked, answers with a 4xx, or its file is too large: send another URL), URL_UNREACHABLE (the host failed: retry later), or INVALID_FILE_TYPE (the file is not a PDF). Returns: {submission_id, status, source_url, accepted_at}. submission_id is the brand_submission_id to pass to the other tools.

extract_photography

ChatGPT
Profile what a brand's photos look like, from the URLs of 1 to 12 photos. Each photo is read against a fixed schema and its pixels are measured, then the photos are merged into one profile: the rules (a value at least 60% of the photos share), what varies, and tone measured from pixels. Use it to brief or check new photography so it looks like the brand's. Starts an asynchronous job and returns immediately. Typical flow: 1. extract_photography(urls) -> returns a brand_submission_id. 2. poll_brand_extraction(brand_submission_id) -> poll until status is "completed" (or "failed"). 3. get_photography_result(brand_submission_id) -> read the profile. Costs 2 credits, refunded when no profile comes back. A photo that is a graphic, a render, or a composite is read, then left out of the profile. A failed submission's error starts with NOT_ENOUGH_PHOTOS (send other photos) or PHOTO_READS_FAILED (retry as is). Returns: {submission_id, status, source_url, accepted_at}. submission_id is the brand_submission_id to pass to the other tools.

extract_website

ChatGPT
Extract the brand system of a website from its URL. Use the result to replicate a site, build new on-brand pages or whole websites that follow the brand, or find visually similar brands. Starts an asynchronous extraction job and returns immediately. Typical flow: 1. extract_website(url) -> returns a brand_submission_id. 2. poll_brand_extraction(brand_submission_id) -> poll until status is "completed" (or "failed"). 3. get_website_result(brand_submission_id) -> read the brand system. Pass sections=["colors", "typography"] to extract only those sections. Pass sections=["brand_assets"] for hosted, judged campaign images and semantic labels from external sources. Enabled full extractions include this section. Dependencies still run when needed; layout-dependent selections may take similar time to a full extraction. Omit sections to extract the full brand, which is required for search_similar_brands. Caching: a recent full extraction of the same URL can be reused, including for a selective request (only requested sections are returned). Selective results are stored for retrieval but never used as cache sources. poll_brand_extraction reports cache_hit: true on reuse. A cache hit is faster but NOT free -- every extract_website consumes 2 credits, cache hit or not. Pass force=true to force a fresh re-extraction (also billed). To re-read a result you have already paid for without spending again, call get_website_result with its existing brand_submission_id (or find it via list_brand_extractions) instead of submitting the URL again. Local pages: a server on localhost is not reachable from the Taste Engine. Expose it over the Tastelabs SSH tunnel (steps in the server instructions) and pass the printed http://crawl-<uuid> URL. Tunnel aliases are never cached, so each call re-extracts whatever is published behind the alias. Returns: {submission_id, status, source_url, accepted_at}. submission_id is the brand_submission_id to pass to the other tools. status begins as "accepted" and advances through queued/crawling/extracting.

get_brand_adherence_result

ChatGPT
Fetch the verdict of a brand-adherence job: the score, the recommendations, and the fixes. Call this once poll_brand_adherence reports status: "completed". Mid-run it can return a partial verdict (components the run has not produced yet stay null or empty, and status reports how far along the job is); early in a run it returns an error telling you to keep polling instead. Reading the verdict: - score: one 0-1 blend of the judge and the deterministic checks, published in 0.05 steps. The judge scores identical inputs up to 0.08 apart; treat a difference smaller than that as variance, not as a real change. - recommendations: the judge's recommendation strings, worst-first, with exact target values to apply. - fixes: the deterministic verifier's fix objects, worst-first, capped at 20.

get_brand_asset_result

ChatGPT
Fetch the design system of a completed brand asset extraction. Call this once poll_brand_extraction reports status: "completed". Before that it returns an error telling you to keep polling. Top-level response: id, status, source_url, medium (billboard, poster, ig_story, ig_post, social_post, or google_ad), and signed image URLs: original (as sent) and flat (the creative, straightened and cropped). result is the design system: medium, format, profile, colors, typography, surfaces, elements, layout, assets (the images cut from the piece and where they sit), not_measured (what could not be read), and actions or icons when the piece has them.

get_brand_extraction_result

ChatGPT
Fetch the extracted brand system for a completed extraction. Call this only once poll_brand_extraction reports status: "completed". If the extraction is still running it returns an error telling you to keep polling poll_brand_extraction, since the result is not ready yet. Prefer sections to keep responses small: the full brand system is often 10-30k tokens, so request only the sections you will actually read (e.g. sections=["colors", "typography"] for palette + type). Omit sections to return all originally requested sections. This filter only narrows the stored result; it cannot add sections omitted from extract_brand.sections. Top-level response: submission metadata (submission_id, extraction_id, status, source_url, cache_hit, completed_at, sections, ...) plus result, which has result.artifacts and result.design_system (the brand system). Top-level sections records the original selection, or null for a full extraction. result.artifacts: captured files, each {url, path, content_type}: html, css, screenshot, full_page_screenshot (any may be null). For replication or visual understanding, fetch full_page_screenshot.url (or screenshot.url) to see the rendered page and html/css for the exact markup. result.design_system is the brand system. Its sections are the values accepted by sections; optional ones (data_display, structure, icons) appear only when detected.

get_brand_kit_result

ChatGPT
Fetch a completed brand kit. Call this once poll_brand_extraction reports status: "completed". Before that it returns an error telling you to keep polling. Prefer sections, and request only what the task needs: social posts and ads colors, typography, logo, photography, graphics, rules web pages and UI colors, typography, buttons, cards, layout, rules copy profile, rules motion and video motion, colors, logo, rules Top-level response: submission metadata (submission_id, extraction_id, status, source_url, completed_at, ...) plus result. result.design_system is the kit, with kind: "brand_kit". result.artifacts has brand_kit_original (the PDF), brand_kit_pages (one rendered image per page), and screenshot (the first page). Fields that help most: colors.palette[].proportion (how much of each colour to use), logo.misuse and the rules whose kind is "dont" (what the book forbids), buttons[].css and cards[].css (drop-in styles), and each asset.url (real logos, graphics and photos from the book).

get_photography_result

ChatGPT
Fetch a completed photography profile and what became of each photo. Call this once poll_brand_extraction reports status: "completed". Before that it returns an error telling you to keep polling. Top-level response: id, status, photos (how many were sent), profile, and outcomes. profile has rules (values at least 60% of the photos share), varies (values no rule covers), tone (exposure, highlights, shadows, contrast and grain, measured from pixels), signature, and palette (the photos' dominant colours, not brand colours). outcomes has one entry per photo: status (read, invalid, or failed), main_subject, excluded (why a read photo was left out of the merge), and error.

get_search_result

ChatGPT
Read a past search and its stored result cards without spending credits. Returns search metadata plus results, the same cards returned by the original search. Works for both search and similar-search history entries. results is null when no cards were stored (a run that has not completed, a failed run, or a search predating result storage), and [] when a completed search matched nothing. Only the owner can read the search; missing or inaccessible submissions return not found.

get_website_result

ChatGPT
Fetch the brand system of a completed website extraction. Call this only once poll_brand_extraction reports status: "completed". If the extraction is still running it returns an error telling you to keep polling poll_brand_extraction, since the result is not ready yet. Prefer sections to keep responses small: the full brand system is often 10-30k tokens, so request only the sections you will actually read (e.g. sections=["colors", "typography"] for palette + type). Omit sections to return all originally requested sections. This filter only narrows the stored result; it cannot add sections omitted from extract_website.sections. Top-level response: submission metadata (submission_id, extraction_id, status, source_url, cache_hit, completed_at, sections, ...) plus result, which has result.artifacts and result.design_system (the brand system). Top-level sections records the original selection, or null for a full extraction. result.artifacts: captured files, each {url, path, content_type}: html, css, screenshot, full_page_screenshot (any may be null). For replication or visual understanding, fetch full_page_screenshot.url (or screenshot.url) to see the rendered page and html/css for the exact markup. result.design_system is the brand system. Its sections are the values accepted by sections; optional ones (data_display, structure, icons) appear only when detected.

list_brand_adherence_jobs

ChatGPT
List the caller's brand-adherence jobs, most recent first. Use this to find an existing adherence_job_id (the job_id field on each item, e.g. to re-read a verdict without paying again) or to review verification history. Narrow with status; paginate via limit/offset. Returns: {jobs: [...]}. Each item has the same fields as poll_brand_adherence.

list_brand_extractions

ChatGPT
List the caller's extractions, most recent first (deduplicated by URL + pipeline). Use this to find an existing brand_submission_id (the submission_id field on each item, e.g. to re-read a result) or to review extraction history. Narrow with status and/or search (URL substring); paginate via limit/offset. Returns: {submissions: [...], total, totals: {status: count}, limit, offset}. Each item in submissions has the same fields as poll_brand_extraction.

lookup_slop

ChatGPT
Look up the design dimensions closest to a described design choice. Searches a design rubric where each dimension pairs a recognizable AI-generated-design anti-pattern (slop_excerpt) with what good and great look like on the same dimension. Use it while designing or reviewing UI to check whether a choice is an AI-slop tell and what to do instead. Synchronous. Free — costs no credits.

poll_brand_adherence

ChatGPT
Check an adherence job's live status and progress. Lightweight; use this to poll. Poll after verify_brand_adherence until status is "completed", then call get_brand_adherence_result. Stop on "failed" and read error for the reason. In-progress statuses: "accepted", "extracting", "judging". Returns: {job_id, status, current_step, reference_url, candidate_url, cache_hit, is_public, accepted_at, completed_at, error}. current_step names the pipeline stage in progress (e.g. "waiting_for_extractions", "judging_scene"). Does NOT include the verdict; use get_brand_adherence_result for that.

poll_brand_extraction

ChatGPT
Check an extraction's live status and progress. Lightweight; use this to poll. Poll after extract_brand until status is "completed", then call get_brand_extraction_result. Stop on "failed" and read error for the reason. In-progress statuses: "accepted", "queued", "crawling", "extracting". Returns: {submission_id, status, current_step, source_url, cache_hit, extraction_id, credits_consumed, accepted_at, completed_at, error}. current_step names the pipeline stage in progress (e.g. "crawling", "design_extract_ai"). Does NOT include the brand system; use get_brand_extraction_result for that.

search_brands

ChatGPT
Search the brand corpus by aesthetic, described in natural language. Runs synchronously and returns ranked result cards, no polling. Each card has {url, brand_name, identity_paragraph, tags, palette, typography, screenshot_url}; fetch the screenshot url to see the brand. Results come from the curated corpus, not from your own extractions; to work with one (replicate it, build on-brand pages), extract it via extract_brand(url). page_type, industry, hue, and layout are optional hard filters, each matched against the fixed set of values listed on its own parameter: every result must satisfy every filter given, on top of the query's own aesthetic match. A value outside that set is rejected as invalid input before the search runs (no credit charged). Costs 1 credit (fast) or 2 credits (deep), refunded if the search fails. Keep the returned search_id: pass it to get_search_result to read these results again for free.

search_similar_brands

ChatGPT
Find brands in the corpus visually similar to one of your extractions. Give it the brand_submission_id of a completed full extraction (extract the site with extract_brand, omitting sections, first if you haven't). Selective extractions are rejected. It embeds that brand system and returns its nearest neighbours as the same result cards as search_brands. Only neighbours close enough to count as similar are returned, so you may get fewer than top_k — an empty list means nothing in the corpus cleared the similarity floor for this brand. If the extraction's visual description identifies an error, verification or blank page without a usable brand capture, the call fails and tells you so: re-run extract_brand(url, force=true) and try again. Synchronous, no polling. Costs 1 credit, refunded if the search fails, is rejected, or returns nothing. Returns {search_id, results}. Pass search_id to get_search_result to read these results again for free.

verify_brand_adherence

ChatGPT
Judge how well a page adheres to a reference site's brand. Use it to score a rebuild, a generated page, or a redesign against the brand it is supposed to follow. Starts an asynchronous brand-adherence job and returns immediately. Typical flow: 1. verify_brand_adherence(reference_url, candidate_url) -> returns an adherence_job_id (the response's job_id field). 2. poll_brand_adherence(adherence_job_id) -> poll until status is "completed" (or "failed"). 3. get_brand_adherence_result(adherence_job_id) -> read the verdict. Both URLs are extracted automatically and a recent extraction of either side is reused, so extract_brand is not a prerequisite. Costs 2 credits per job (covering both sides, cache hit or not), refunded if the job fails. To re-read a verdict you already paid for, call get_brand_adherence_result with its existing adherence_job_id (or find it via list_brand_adherence_jobs) instead of submitting the pair again. Local pages: a page served on localhost is not reachable from the Taste Engine. Expose it over the Tastelabs SSH tunnel (steps in the server instructions) and pass the printed http://crawl-<uuid> URL as candidate_url. Returns: {job_id, status, reference_url, candidate_url, accepted_at}. job_id is the adherence_job_id to pass to the other tools. status begins as "accepted" and advances through extracting/judging.

Capabilities

Writes

Example Prompts

Click any prompt to copy it.

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/brand-api"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/brand-api)