Prism

A Prism endpoint is a composite: one call fans out to several platforms in parallel, merges the results, and returns a single computed answer. You do not orchestrate the fan-out, the retries, or the merge yourself.

Base URL: /v1/prism/...

What a composite is

A normal endpoint reads one thing from one source. A composite answers a question that no single source can answer, by calling several of them and doing the reconciliation work in between.

Three properties follow from that, and all three change how you write the client:

1. A composite has legs, and a leg can fail without failing your call. Responses carry a coverage signal (legs[], sources_succeeded, coverage, or a confidence block, depending on the composite) telling you which sources answered. Read it before you read the numbers.
2. A composite is priced per call, not per leg. The price does not go up because a source was slow or a fan-out went wide, except where the request shape itself sets the width.
3. A composite returns the same unified schema as its parts. A profile inside a composite response is byte-identical to the same profile from its native endpoint, so one parser handles both.

When to use a composite instead of N calls

Reach for a composite when the answer needs cross-source reconciliation and you would otherwise write it yourself: deduplication across platforms, weighting one source against another, matching the same entity under two different ids, or holding a fan-out open while slow legs settle.

Stay on the per-platform endpoints when you want one source's raw data, when you need a field the composite does not surface, or when you are refreshing something you already hold. prism/lookup and prism/post-stats exist precisely for that last case and are priced accordingly.

How Prism is priced

Composites come in three pricing shapes.

The bulk endpoints are metered per item at each row's own platform rate, so a 50-handle profiles batch of TikTok handles is 50 credits while the same batch of LinkedIn handles is 250. Only successful rows are charged.

Quote ranges rather than floors when you budget: the roster below shows each composite's base tier price, which is the bottom of the range for the metered and param-derived ones.

For handle-audit, a public response-cache hit costs 0 credits. A non-200 execution failure after deduction receives a full refund, and request-validation errors happen before billing. If every leg fails, the whole call is refunded. Every successful HTTP 200 receives the full request-shaped charge even with partial evidence, and there is no partial refund on that path.

Quickstart

curl "https://www.socialcrawl.dev/v1/prism/handle-audit?handle=mrbeast" \
  -H "x-api-key: YOUR_API_KEY"

4. Read computed fields

Direct post and author endpoints inside a composite response may include the shared computed fields such as engagement_rate, language, and content_category. See Computed fields for formulas and null semantics.

Brand health

How is this brand doing right now?

share-of-voice is 20 credits per brand web-only and 40 per brand with social (the default), so the price is fixed by how many brands you pass and whether include contains social. Pass competitor to earned-media to see which outlets covered them and not you.

A brand name is often also a word, a song or another company. brand-mentions, share-of-voice, earned-media, campaign and crisis-radar judge whether each row is about your brand and publish cleaned numbers next to the original ones, never in place of them: sov_pct_about_brand beside sov_pct, clipping_count_about_brand and top_outlets_about_brand on earned-media, top_amplifiers_about_brand on campaign, score_about_brand and grade_about_brand on crisis-radar. Add brand_description (one plain sentence, such as "Samyang's spicy instant noodle brand") when the name is ambiguous. Prices are unchanged.

# Is the conversation moving against us, and by how much?
curl "https://www.socialcrawl.dev/v1/prism/crisis-radar?brand=Figma" \
  -H "x-api-key: YOUR_API_KEY"

Creators

Should we work with this person?

handle-audit and creator-card are 5 credits for up to 4 platforms and +1 per extra. Run handle-audit before you spend on the pull.

# 1. Vet the handle and get a cost projection before committing
curl "https://www.socialcrawl.dev/v1/prism/handle-audit?handle=mkbhd" \
  -H "x-api-key: YOUR_API_KEY"

# 2. If it scores well, run the full partnership check
curl "https://www.socialcrawl.dev/v1/prism/creator-vet?handle=mkbhd&platform=youtube" \
  -H "x-api-key: YOUR_API_KEY"

Commerce and apps

What do buyers actually say?

product-reviews tells you what buyers said. review-integrity tells you whether to trust the stars. app-reviews covers review text. apps-lookup compares the listings, and its title-match guard is what stops two apps sharing a name from being merged.

When you pass a name instead of an id, these composites (and reputation) look at the top few search results on each source and judge which one is the product, business or app you meant, instead of taking the first hit, so an accessory, a reseller or a same-name app is no longer graded as yours. The choice is reported per source in entity_resolution (resolved_by, index, title, confidence), and when no result is the right one the source is left out with entity_not_found:<source> in _warnings. Pass asin, gid or a store id to skip the judgment.

# Reviews for one product across marketplaces
curl "https://www.socialcrawl.dev/v1/prism/product-reviews?asin=B0DGHMNQ5Z" \
  -H "x-api-key: YOUR_API_KEY"

Developer ecosystem

How is this tool landing with engineers?

launch-echo is the single-moment view; devtool-pulse is the standing one. org-radar is 1 credit base plus 5 per repo expanded, up to 10 repos. Set repos to control it.

curl "https://www.socialcrawl.dev/v1/prism/org-radar?org=vercel&repos=3" \
  -H "x-api-key: YOUR_API_KEY"

AI visibility and research

A probe on ai-visibility is one prompt times one run times one engine, so a 5-prompt, 3-run, 2-engine sweep is 60 credits. The ceiling across the widest possible sweep is 1605, so set prompts, runs and engines deliberately, or use preset=quick|standard|deep, which sets runs and caps prompts for a flat probe budget. The engines are Perplexity Sonar and Grok, and appearance is detected deterministically via recognition tokens rather than by a model's judgement. Add include=web_baseline to see which AI-cited domains you do not yet rank on. No per-platform endpoint covers this.

audience-questions judges every title and comment for whether it asks a real question, so questions written without a "?" (common in Korean) are kept and rhetorical ones are not. methodology.question_detection reports how many the judgment added and removed against the old question-mark rule.

# Start narrow: 1 engine, 1 run, a short prompt list
curl "https://www.socialcrawl.dev/v1/prism/ai-visibility?brand=SocialCrawl&topic=social%20media%20api&engines=perplexity&runs=1" \
  -H "x-api-key: YOUR_API_KEY"

Go-to-market, hiring, and Korea

Truth Social has no search, so truthsocial-pulse is the way to watch an account without chaining calls. On korea-gap the social leg is the default. Drop social from include if you only need the web comparison.

Buyer leads from comments

prism/comment-leads finds the people who want what you sell in the comments under a keyword search, and opens only the posts worth opening, because fetching and reading every post's comments is where the cost goes.

curl "https://www.socialcrawl.dev/v1/prism/comment-leads?query=standing+desk+setup&offer=Height-adjustable+standing+desks+delivered+to+your+home" \
  -H "x-api-key: YOUR_API_KEY"

It searches TikTok for query (TikTok only for now), judges each post that reports comments for whether people who want your offer would plausibly comment under it, skips the clearly off-offer ones, and opens the busiest of the rest, up to open (1 to 20, default 12). Each fetched comment is labelled for purchase intent, for being a real question, for spam, and for whether your offer would help its author. A comment is a lead when purchase intent is at least 0.5, spam is under 0.5 and offer fit is at least 0.5. include=askers adds comments that ask a question your offer answers.

Each entry of leads[] carries the comment id and text, the commenter's public username and display name as the platform shows them, the post it came from and every probability. posts_opened and skipped_posts say which posts were opened and why the others were not (no_comments, no_url, below_gate, over_open_limit). economics shows the bill: posts searched and opened, comments fetched and judged, credits by leg, credits.held and credits.refunded, naive_credits_estimate (what opening every post would have cost at the same prices), credits_saved_vs_naive and credits_per_lead.

You pay for what is opened: 1 credit a search page, 1 a comment page fetched, and 1 for every started 25 comments labelled. Judging which posts to open is free. The default request holds 37 credits and refunds what did not run. Most of the saving comes from opening the busiest posts first; the judgment mainly skips posts that are clearly off-offer. Leads are labels for your own review, and nothing is contacted.

Background screening

prism/adverse-screen reads the recent public posts of a person's own profiles and points a reviewer at the posts that may be adverse, each with a citation. It is built for background-screening and KYC teams.

curl "https://www.socialcrawl.dev/v1/prism/adverse-screen?twitter=elonmusk&name=Elon+Musk&categories=fraud_or_financial_crime,violence_or_threats" \
  -H "x-api-key: YOUR_API_KEY"

Identity first. With name, each profile is first checked against the named person, and only profiles judged to be their own are screened. The rest are listed in profiles with the verdict and are not screened, so a namesake's posts are never attributed to the person. A profile judged to be someone else is never screened, even with screen_unconfirmed=true. Without name there is no identity check and you vouch for the profiles.

What it judges. For each post and each category you ask for: does the author engage in, admit, threaten or promote it. A separate check keeps posts that report, quote, warn about or condemn what someone else did (news, a scam warning) from counting as the author's conduct. Every flag in flags[] cites its post (platform, post_id, url, published_at, an excerpt) with the probability per category and a severity from 1 (minor) to 3 (severe). category_counts counts flagged and uncertain posts, and requires_review is true when anything is flagged.

What it never does. There is no summary, no verdict and no score about the person, and automated_decision is always false: a flag is a pointer to a post for a human reviewer to read, and a flag alone must never reject a candidate. Political opinion, religion, health, sexuality, gender, ethnicity, nationality, age and disability are never screened or inferred. It reads public posts only; private accounts return nothing. It reads post text, not images, video or audio. sanctions_or_terror_links and sexual_misconduct had no real positives in the measurement set, so they are unmeasured; see how we measure.

It costs 25 credits for one profile plus 10 for each further profile (4 profiles cost 55), refunded down to the profiles actually screened. A call that found profiles but screened none keeps 5. screen_complete is false when a profile or a post could not be screened, and _warnings says why.

Per-object workhorses

The cheap ones you call constantly rather than once a week.

lookup sniffs the platform and resource, dispatches, and adds no surcharge: a social post routes to 1 credit, an Amazon product to 5. It takes ITEM links only, so a profile, channel or storefront URL is a free 400 pointing you at that platform's own /profile endpoint. The resolved block names the platform, endpoint, and archetype so an agent knows the shape that follows.

post-stats bills 1 credit on most platforms, 2 for an Instagram post and 5 for LinkedIn; profiles bills 1 credit on most platforms and 5 for LinkedIn. Dead, errored, and unsupported URLs are refunded and never fail a post-stats batch, and only successful rows are charged on profiles. profiles supports instagram, tiktok, youtube, twitter, threads, twitch, snapchat, truthsocial, bluesky, kwai, linkedin, and facebook. That roster is this endpoint's own, and is unrelated to the universal-search fan-out. With "include": "posts" (up to 25 items) each ok row also carries a posts block with that handle's first page of posts on Instagram, TikTok, YouTube, X, Threads and Facebook, for 1 credit more per row whose page came back with posts; since and a per-item stop_at_id stop each page where your last check ended. See Batch endpoints. For more than 50 handles or post URLs, POST /v1/prism/jobs runs up to 5,000 profiles or post-stats items in the background at the same per-row prices; see Background jobs.

comment-lookup takes each item as either { comment_url } or { platform, post_url, comment_id }, bills TikTok at 2 and Instagram at 5 (more with deep_scan), and holds at most 100 credits for the whole batch. Store each row's lookup.position_hint and pass it back next run to keep a daily refresh near the floor.

# One URL, whatever platform it is
curl "https://www.socialcrawl.dev/v1/prism/lookup?url=https://www.youtube.com/watch?v=dQw4w9WgXcQ" \
  -H "x-api-key: YOUR_API_KEY"

# Refresh engagement on a batch of posts
curl -X POST "https://www.socialcrawl.dev/v1/prism/post-stats" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "content-type: application/json" \
  -d '{"urls":["https://www.tiktok.com/@nasa/video/7259262686435118378","https://www.youtube.com/watch?v=dQw4w9WgXcQ"]}'

All endpoints

51 endpoints available.

Platform notes

• Composites use GET with query parameters, except the bulk endpoints (post-stats, profiles, comment-lookup, post-comments), which take a POST JSON body.
• Several composites take an include CSV that switches optional legs on or off. On share-of-voice, korea-gap, creator-vet and video-intel that CSV also moves the price, so read it as a cost control rather than a display flag.
• Long-running composites can stream Server-Sent Events when you send Accept: text/event-stream, emitting each leg as it settles. This is the right mode for the 30-credit-and-up composites, which do real fan-out work before they answer.
• A leg can fail without failing your call. Read the coverage signal before the numbers.
• Responses follow the unified SocialCrawl schema, so a profile returned by a composite is byte-identical to the same profile from its native endpoint.

Next steps