One endpoint, 1,585 product videos, every one scored. Ask for a row and
get the hook, the numbers, the written read and the shots, 37,736 cuts
timed from the files themselves. Nothing is invented: a field that can't be judged
from pixels stays null.
49tools
699formats
37,736timed shots
1,585videos
Connect
Streamable HTTP, JSON-RPC 2.0. Send a member bearer token, or an x-sauce-license key for the organic tools. Reading is free: initialize, tools/list, the seven organic tools and the three workflow tools answer without a key. The shelf tools still need a member token.
POST https://sauce-ops.appskins-studio.workers.dev/api/aisle/mcp Authorization: Bearer <member_token> x-sauce-license: <licence_key> Content-Type: application/json
Handshake: send {"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}, then tools/list, then tools/call with {"name":"teardown_get","arguments":{"id":282}}. The full schema for all 54 tools comes back from tools/list; the cards below are the common calls.
What a teardown actually contains.
One teardown per video, one evidence record per shot. The editorial read is stored; the timings come from the file; the frame read is used to understand the video. If a field cannot be judged from pixels, it stays null.
score · score_hook · score_clarity · score_persuasionSwipe score out of 100 plus the three subscores.
awareness · hook · captionThe written read: awareness level, spoken hook and the caption.
views · likes · comments · savesEngagement stored with the row.
frame_urlThe sampled still for that shot, served by /api/aisle/scene.
shot_sizeclose up, medium, wide, read from the frame.
camera_angleeye level, high, low, overhead, read from the frame.
settingWhere the shot happens, e.g. kitchen counter.
subjectThe main thing in frame.
product_visible · product_descriptionWhether the product is in frame and what it looks like.
actionWhat is happening inside the shot.
on_screen_textOCR of the frame, verbatim.
lighting · dominant_colourRead from pixels.
transition_inCut type from the manifest: start, scene, diff or slide.
made_with · made_with_evidencelive-action, synthetic or unknown, only with a specific pixel artifact.
camera_moveAlways null. One still per shot cannot prove movement.
why_it_worked · retention_beats · stealable_linesThe written read: why it worked and the lines to reuse.
camera_move is deliberately null. A single still can prove composition, not motion. The editorial fields are labelled as stored teardown data, not as pixel measurements.
Organic winners
These find what is still running now, not what once ran. One input: a product URL, a competitor store, a post link, or a plain niche. Reading winners is free. One full brief per niche is free; the pack costs one run against a licence key.
resolve_input
Any input to a niche
Resolves a product URL, store, post link or plain query into product, category, audience, market and language. Ask the user once if the guess is wrong, then pass market and language back.
keys input, market, language
{"input":"posture corrector for tall men","market":"US","language":"en"}
search_organic_winners
Current winners, deduped
Videos, slideshows, native posts and UGC, collapsed to one creative per concept with a post count. Each row carries the still-running score /100, status, profit label, confidence and evidence. Filter by platform, format, status, score, product, reposts and age.
{"input":"posture corrector for tall men","platform":"tiktok","min_ows":60,"limit":40}
get_teardown
One winner, timed
First frame, hook, every beat with seconds, proof, CTA, pacing, format, on-screen text, sound, creator and product, plus the shoot-ready brief for that creative.
keys input, creative_id, market, language
{"input":"posture corrector for tall men","creative_id":"cr_f0d31caab17e"}
explain_score
Why it ranks
The six weighted components, the confidence and status labels, the evidence fields and the limits. Nothing is a black box; the persistence proxy is labelled as a proxy.
keys input, creative_id
{"input":"posture corrector for tall men","creative_id":"cr_f0d31caab17e"}
build_brief
The niche pack
Ten hooks, three scripts, a 14-shot list, b-roll, overlays, creator direction, sound, exports, posting plan, test matrix and falsifiers, plus Markdown. One full brief per niche is free; the pack costs one run.
keys input, creative_ids, market, language, format
{"input":"posture corrector for tall men","creative_ids":["cr_a","cr_b","cr_c"],"format":"markdown"}
watch_niche
Hand the watching to your AI
Stores a baseline and returns a watch token. AISLE sends no alerts: your agent polls diff_since on its own schedule and decides who to tell.
keys input, market, language
{"input":"posture corrector for tall men","market":"US","language":"en"}
diff_since
What changed
New, climbed, dropped and dead against the stored baseline, with a fresh brief for the best new creative. The token is scoped to the caller.
keys watch_token
{"watch_token":"w_a4cee18d3afac5b0"}
Workflows one per buying moment.
Fifteen agent workflows, each with the full prompt, the ordered MCP calls, the output contract and the cases where the agent must stop. The same record is a JSON payload, a page at /aisle/workflows, and a SKILL.md file. Paid-media tooling is one workflow, labelled BOARD - it never appears in the organic moments.
workflow_list
All fourteen workflows
Recreate, source, validate, buy, contracts, play-stack, agent, full stack, launch, shoot, write, clip, supplier, produce. The write and clip runbooks are free, no key. Each row carries the agent role, the inputs, the MCP calls and the output shape.
keys q, intent
{"intent":"validate"}
workflow_get
The full runnable record
Mission, inputs, every step with its exact call, the complete agent prompt, the output contract, the checks and the refusal cases. Paste the prompt into Claude, Codex or your own runner.
keys id
{"id":"recreate"}
skill_get
The same record as SKILL.md
Frontmatter, mission, inputs, run steps, output contract, rules and refusals. Save it to a skills folder or download it from /aisle/skills/recreate.md.
keys id
{"id":"recreate"}
Cut and frame tools
These read the scene manifests, the stored pixel reads and the library counters. Payloads are complete: paste them as the arguments object.
shelf_search
Search the shelf
Free text over hook, caption, creator, format, platform and pixel summary. Each row returns its score, awareness, engagement, editorial read, cut state and shot count.
keys q, category, format, min_views, sort, limit, offset
Score with hook/clarity/persuasion subscores, awareness, hook, caption, engagement, why it worked and reusable lines, plus the manifest state, shot count, pixel summary and production verdict.
keys id
{"id":282}
teardown_get
The teardown and its evidence
The full editorial read plus real cuts, the frame table, the stored per-shot pixel read and the production verdict. Editorial and frame fields are returned as separate blocks.
keys id
{"id":282}
teardown_search
Search the teardowns
Free text over hook, caption, creator, structure and the stored pixel summaries. Each result returns its score, awareness, editorial read and cut state.
keys q, limit
{"q":"restock","limit":10}
reverse_video
Cut a real video scene by scene
Real cut timestamps from ffmpeg scene detection and frame diffs, per-shot frame URLs, a labelled contact sheet, a vision prompt and a skill prompt. Returns status:"pending" instead of inventing cuts.
keys row_id, id
{"row_id":282}
describe_scenes
Make a vision model read the frames
Sends the real frame images to the caller's own vision key and returns per-shot grammar plus made_with and the pixel evidence. This is how the tool sees the video; the description fields are evidence for the teardown, not the teardown itself.
shots and frame strip, first-frame hook, first product appearance, pacing (cuts, average shot, shortest and longest), CTA from on-screen text, production verdict and provenance.
keys row_id, id, format
{"row_id":282}
recreation_brief
Hand a video to a model
A machine payload: shot plan with frame URLs and timings, per-shot pixel fields, product context when the shelf has it, paste-ready prompts and an agent prompt for Claude or Codex.
keys row_id, id, url, format
{"row_id":282}
swipe_export
Take the teardown with you
Markdown or CSV: id, url, platform, creator, swipe score with hook/clarity/persuasion subscores, awareness, hook, caption, engagement, manifest state, shot count, described shots and the pixel summary.
keys q, category, format, min_views, limit, out
{"min_views":300000,"limit":40,"out":"csv"}
stats_overview
Library counters
Videos, cut manifests, timed videos, photo carousels, shots, median shot, under-2s share and unavailable count.
keys none
{}
shelf_signals
Distributions
Shot-length buckets, shots per video, cut types and platform counts, computed directly from the manifests.
keys none
{}
creator_list
Find the creators worth studying
Ranks creators by total views, with video counts and best video by views. Filter by handle.
keys q, limit
{"q":"restock","limit":20}
creator_get
Study one creator
Video count, total and median views, best video by views, and their top five rows with pixel summaries.
keys handle
{"handle":"sam.shan.shops"}
analyze_video
Tear down one video with your key
Sends the shots and the stored pixel read only. The caption, handle, bio, audio and views are withheld. Returns what the frames show and what to copy.
keys row_id, provider
{"row_id":282,"provider":"openai"}
Creative tools write new copy.
These generate or grade text for a product. They do not judge a video, and they never invent a view count.
make_scripts
Three shootable scripts
Hook, timed beats, proof shot, on-screen text, CTA, shoot notes and a citation to a shelf video.
Reads the row's format and category, then writes fresh hooks for your product and audience.
keys row_id, audience, count
{"row_id":282,"audience":"home cooks","count":15}
score_draft
Grade a draft you wrote
Hook, clarity, proof, structure and CTA out of 20 each, the weakest dimension and exact fixes. It grades text; it never claims to judge a video.
keys script, product, hook
{"script":"I burned eggs three mornings in a row. Then I switched one thing...","product":"ceramic pan"}
score_batch
Grade a week of drafts at once
Up to 20 at once: each gets its total, grade, dimensions and the exact fixes. Text only.
keys scripts, product
{"scripts":[{"name":"A","script":"I burned eggs three mornings in a row. Then I switched one thing..."},{"name":"B","script":"Stop scrubbing your pan. Here is the 40-second test."}],"product":"ceramic pan"}
Library and retrieval
The format library and the closest videos, with the numbers that actually exist.
similar_winners
Closest videos for a product
Ranks the shelf by category, format, price proximity and views, then returns the rows with their numbers and cut state.