Heista MCP server

Get the current status of a specialist dispatch job started via dispatch_<specialist>_async. Returns { status: queued|running|completed|failed, result_text?, error_text?, error_class?, retry_count, elapsed_seconds, wait_ms_hint }. Call this repeatedly after a dispatch_*_async returns a job_id. Sleep wait_ms_hint milliseconds between calls. When status === "completed", read result_text as the specialist's full synthesis. When status === "failed", error_class tells you whether to retry (transient/scope/routing) or give up and synthesize around (permanent) per the fleet resilience pattern.

Read-only walk of a fleet session tree. Given any session_id in the tree (root, Head, Mastermind, or specialist sub-node) returns the full breakdown: every session row with depth + parent + agent_kind + node_label, the cost_events recorded against each, per-node self_cost_cents, total raw compute, tier markup estimate, and (after close_session_tree has run) the authoritative credits_charged + credits_refunded. Org-scoped: only sessions belonging to your org return data. Free — no compute cost. Use to render cost breakdown UIs, audit fleet spend, or verify a session's tree topology.

List saved assets in the workspace. Filter by category (STRATEGY, IDEAS, COPY, VISUALS, MOTION, BRIEFS), by one or more formats inside the category (e.g. COPY + formats=["ad-script","hook"]), by tags (any/all), by brand_id, by brief_id (PowerSource), by created_by ("me" resolves to caller via OAuth), or favorites_only. Returns the unified view that backs the /assets page — BRIEFS rows come from creator_briefs with share URLs; other categories come from saved_assets. Use BEFORE asking the user what to pull into a Heist. Free, read-only, paginated.

Fetch one saved asset by id. Returns the full row including category, format, tags, body_text/html, signed media_url (if private storage), metadata, creator, brand, and timestamps. Use AFTER list_saved_assets to load the full record when the list projection is too sparse.

Fetch up to 50 saved assets by id in one round-trip. Use when an agent needs to pull a pre-selected set — e.g. resolving a saved_asset_picker context input on a Heist that requires N pinned assets. Missing or cross-workspace ids are silently dropped; compare returned items vs requested ids to detect drops.

Persist a new saved asset to the workspace. category MUST be one of STRATEGY, IDEAS, COPY, VISUALS, MOTION (BRIEFS lives in creator_briefs and is not saveable through this tool). format MUST match the per-category enum (see input description). title is required. body_text + metadata are recommended. Source attribution (heist_slug, session_id, pattern) lets the user trace the save back to its origin in the /assets timeline. Brand and brief_id link the save to a PowerSource for downstream filtering. Returns the inserted asset row.

Delete one saved asset by id. Destructive — confirm with the user before calling. OAuth callers can only delete saves they created themselves (Linear model — see /assets UI for org-admin override). API-key callers are treated as org-trusted and can delete on behalf of any creator in the workspace. Cleans up the storage object for private VISUALS/MOTION saves.

Toggle the favorite flag on a saved asset. Pass is_favorite=true to favorite, false to unfavorite. favorited_at is set/cleared in lockstep so the Favorites tab sorts correctly. Not destructive.

List audience archetypes for a strategy (PowerSource). Returns the Buyer Decoder archetype (source="buyer_profile", one entry max) plus up to 3 offering primary_audience segments (source="primary_audience"). Use this to pick which audience to target before generating copy / scripts / hooks — the UI picker reads the same projection. Distinct from list_strategies (which lists scans for a brand): this lists audiences INSIDE one strategy.

Build the highest-fidelity creative intelligence profile by combining a brand's public website URL with their internal documents. Takes a required website URL plus at least one document — file_ids from previous upload, public document_urls (PDF/DOCX/TXT/MD, up to 10), or documents_inline (base64-encoded). Optional idempotency_key for safe retry. Returns a job_id; poll with get_powersource. Same response shape as create_powersource_url, but the synthesis cross-checks how the brand presents publicly against what the team actually believes internally, producing stronger conviction on voice, positioning, proof, and tension architecture than either input alone. Use this when the user has both a public site AND a brief / brand guidelines / strategy deck and wants the deepest possible profile — the kind of intelligence a senior strategist produces over a week. Default recommendation when both inputs are available. Costs 200 credits. Do NOT use for URL-only scans — use create_powersource_url (100 credits). Do NOT use for docs-only scans — use create_powersource_docs (100 credits).

Check the calling user's Heista API credit balance, month-to-date usage broken down by operation, lifetime spend, and the current pricing for every paid tool. Takes no inputs. Returns balance in cents, lifetime spend in cents, month-to-date call counts per tool (decode_ad, create_powersource_*, generate_adscript), per-tool unit pricing, and a top-up link the user can follow to add credits. Free, read-only, idempotent. Use this whenever the user asks about credits, balance, usage, how much they've spent, top-ups, pricing, "what does this cost", or "how many credits do I have". This is also the ONLY surface where dollar amounts are legitimate to report in conversation — everywhere else, cost should be referenced in credits, not currency. Do NOT use to add credits or change billing — only to read state. Do NOT call this on every turn — invoke once when the user explicitly asks about account state.

Browse proven hook patterns from Heista's corpus of decoded winning Meta/TikTok ads. Takes optional filters: vertical (e.g. BEAUTY_SKINCARE, SUPPLEMENTS, APPAREL), hook_type (e.g. CURIOSITY_SPIKE, CONTRADICTION, CALLOUT), and marketing_angle. Returns hook examples (the real opener lines from successful ads), pattern templates, the psychological mechanism behind why each one stops the scroll within the first 1.5 seconds, and runtime performance data (active days on Meta when available). Free, read-only, idempotent — no credits consumed. Use this when the user asks "what hooks stop the scroll", "give me hook ideas", "how should I open this ad", "show me hooks for [vertical]", or needs scroll-stopping openers grounded in proven patterns rather than guessed copy. Useful before writing a script — pair with adformula_intelligence or decoder_intelligence for the full beat structure. Do NOT use to decode a specific ad URL — use decode_ad. Do NOT use to generate finished scripts — use generate_adscript. Hooks here are pattern intelligence, not finished copy.

Browse proven ad formula blueprints — structural patterns clustered from 3-10+ winning ads that independently converged on the same beat architecture while Meta kept rewarding them with sustained spend. Takes optional filters: vertical, creative_format (e.g. TALKING_HEAD, UGC, FOUNDER_STORY), marketing_angle, algo_intent, hook_type, and limit (1-10, default 5). Each formula returns: source ad count, average active days (runtime proof), confidence score, 6-layer beat blueprint, per-beat visual direction, marketing angle, psychology mission. Free, read-only, idempotent. Use this when the user asks "what's working in [category]", "show me formulas for talking-head ads", "what scripts work in my vertical", or wants category-level pattern discovery before committing to a single ad. Pass the returned formula id to generate_adscript with source_type="formula" for synthesis. When choosing among results: prioritise (1) avg_active_days as primary proof, (2) marketing_angle alignment with the brand's buyer tension, (3) source_ad_count for cluster robustness, (4) confidence_score as tiebreaker. Do NOT use when the user names a specific ad — decode that ad with decode_ad. Do NOT use for sentence-level transcript fidelity — formulas abstract the structure, not exact copy.

Browse individual decoded ads from Heista's corpus of real winning Meta/TikTok creative. Takes optional filters: vertical, creative_format, marketing_angle, hook_type, algo_intent, brand (partial name match), and limit (1-10, default 5). Each result returns beat timeline, classification, psychology, runtime performance signals (active days on Meta when available), and a decode id you can pass into generate_adscript with source_type="decode" to write a fresh script on that exact structure. Free, read-only, idempotent — no credits consumed. Use this when the user wants a specific ad as a script template (not an averaged formula), asks "show me winning ads in [vertical]", "what are [brand]'s top ads", or wants to see examples before committing to a generation. Source discovery surface — the response is the spine; for the full bundle with transcripts and director's read, call get_decode by id afterwards. Do NOT use to decode a NEW ad from a URL — use decode_ad (paid). Do NOT use for category-level patterns abstracted across multiple ads — use adformula_intelligence. Do NOT use to write the script itself — use generate_adscript or write directly from the bundle.

Generate direct-response video ad scripts by fusing a proven structural source (decoded ad or formula) with a brand's PowerSource. Output is feed-native ad copy for paid social (Meta, TikTok, Reels) in the brand's voice — hook, beat-by-beat body, CTA close, plus visual direction per beat. Takes source_id (from adformula_intelligence, decoder_intelligence, or decode_ad), source_type ("formula" or "decode"), powersource_id (from any create_powersource_*), and tunable params: count (1-5 variants, tensions and selling points auto-rotated across variants), script_mode ("blueprint" preserves source structure exactly, "remix" preserves psychology but writes original copy), duration (target seconds), audience, tension override, selling_points override, voice_mode ("creator" for UGC default, "brand" for owned channels), and idempotency_key. Use this when the user says "write me a script", "I need a TikTok script", "write an ad based on this", or wants shell-faithful replication of a proven winner in their own brand voice. REQUIRES both a structural source AND a powersource — guide the user through creating either if missing. Metered pricing — typically 2-5 credits per script (~2 credits for 15s, ~5 credits for 60s). Pre-flight reserves a 17-credit ceiling and refunds the difference after measurement. Do NOT use to discover sources — use decoder_intelligence or adformula_intelligence first. Do NOT use to extract brand intel — use create_powersource_url first.

Heista's creative direction engine — same engine the Creative Director specialist runs internally, exposed over MCP. ONE-SHOT: give a brief, get N finished creative outputs. For back-and-forth refinement, or output shapes the `medium` enum below does not cover, use chat_with_creative_worlds instead. OUTPUT SHAPE switches on the `medium` arg: • omitted → N territory cards (default exploration). Each card sits on different psychology / craft / feel / world axis coordinates so the set spans the creative space rather than orbiting one insight. Card has: name, campaign line, 5-8 sentence pitch, one-sentence strategic bet, resolved axis state names, creative-director rationale. • `tvc` → N TVC scripts (15-90s — hook, arc, resolve, sound design, end line). • `billboard` / `ooh` / `print` → N out-of-home concepts (visual concept + line + placement rationale). • `social` → N social-video concepts (hook + format type + middle beat + payoff, optimised for Reels / TikTok / Shorts). • `activation` / `experiential` → N activation concepts (space design + user journey + peak moment + takeaway artifact). • `audio` → N sonic / radio concepts (sonic scene + voice + audio arc). • `campaign` → N full campaign platforms (insight → big idea → strategy → visual world → production roadmap). The engine can also produce manifesto / copy, naming, packaging, PR stunts, content series, brand positioning, partnerships — these output shapes are NOT in the medium enum, so use chat_with_creative_worlds when the user wants one of those. USE WHEN: user says "give me ideas / options / directions / territories", "what angles work for...", "show me three / five ways to...", "write a TVC for...", "draft billboard concepts for...", "I need fresh thinking on...". DO NOT USE to refine one existing direction (use chat tool), to critique work, for OKRs / internal docs / strategy decks, or anything outside advertising creative direction. INPUTS: brief (the creative problem, free text), count (2-6 concepts), optional brand_id (from list_brands or any create_powersource_* — when provided the engine grounds output in the brand's buyer tensions, voice, and selling points), optional medium (above), optional lens_hint (apply a playbook or signature move as a creative constraint), idempotency_key (safely retryable for 5 minutes). Returns the finished creative output as narrative text PLUS a structured array of resolved axis coordinates for programmatic use. Metered — typically 3-15 credits per call depending on count and brand context size. Charged after success on actual token usage.

Multi-turn conversation with Heista's creative direction engine — a real chat where the agent decides each turn what to produce based on what you ask for. Use whenever the work needs more than one round, OR when you want an output shape not covered by call_creative_worlds' `medium` enum. WHAT YOU CAN ASK FOR (any of these, turn 1 or any turn after): • Territories — "give me five directions for X", "what angles work here" • A TVC script — "write a 30-second TVC for Cowboys" • Billboard concepts — "three billboards under a quiet-authority lens" • A campaign platform — "build #2 into a full campaign with the big idea" • A manifesto or copy — "draft the manifesto in the brand voice" • Naming — "name this product, five options with rationale" • A PR stunt — "what's the newsworthy version of this" • A content series — "20 episode ideas for a brand podcast" • Packaging, sonic branding, partnerships, social systems • Refinement — "make #2 darker", "extend that into a tagline", "summarise" • Pivots — "forget the soft-drink angle, try the late-night insomnia one" SESSION: omit session_id on turn 1; the response returns a fresh session_id you pass on every subsequent turn — that is how the conversation persists. brand_id is only honoured on turn 1 of a new session (continuing sessions keep their original brand context). USE WHEN: user wants back-and-forth, OR wants an output shape outside the medium enum (manifesto, naming, press release, content series, packaging, etc.). Prefer call_creative_worlds when the user wants "three options, done" with no follow-up. WON'T DO: write OKRs / internal docs / strategy decks; behave as a general assistant. It is a creative director with creative-director taste — anti-cliché, specificity test, will push back on vague briefs. Metered — typically 2-10 credits per turn depending on tool use and context size. Charged after each turn on actual token usage.

List every brand in this workspace. Use this BEFORE creating a PowerSource to avoid creating duplicate brand records (pass the matching brand_id to create_powersource_*), and to discover brands the user can pivot a Heist to. Each row carries the brand_id (persistent identity), name, domain, asset_count, strategy_count, and brand status. Use this when the user asks "what brands do I have", "show me my brands", or before any image-led work where you need to know which brand owns assets. Free, read-only. Distinguish Brand (persistent, brand_id) from PowerSource (a scan, powersource_id). A brand has many PowerSources; pick the brand first, then narrow to a strategy with list_strategies.

Get a brand's full canonical record — name, domain, voice (tone_of_voice), story, visual identity (logo, primary color, visual assets), and counts. Use to inspect what a brand carries before deciding which Heist context to run, or to read the brand voice directly when writing copy. Free, read-only.

List all PowerSource strategies (scans) for a brand. A brand has many strategies — one per scanned URL. Product-page strategies carry product_name and is_product_page=true; use these to label them in conversation or to pick the right one for a product-focused generation. Returns powersource_id (use as the brief/PowerSource id everywhere else), product_name, scanned_at, source_url, is_pinned. Free, read-only. Paginated via cursor.

Dispatch to the MARKET ANALYST — entity-deep teardown of a named brand or vendor. Use for: "what is brand X / how does company Y work / decode competitor Z / teardown vendor W". Multi-axis extraction grounded in multi-class sourcing, plus defensible MOAT and credible GAP theses. Vertical and geography agnostic. Returns: 8-axis extraction (positioning / offer / audience / voice / pricing / distribution / proof / trajectory) + MOAT thesis + GAP thesis + Sources. NOT for: topic landscapes without a named entity (use dispatch_desk_researcher) / trajectory questions about a category (use dispatch_trend_researcher). ASYNC version: returns { job_id } immediately, the specialist runs durably on a Vercel Workflow (no 300s timeout). Use this version when the specialist is expected to take >90s. Call get_dispatch_result(job_id) periodically (respect wait_ms_hint in the response) until status === 'completed' or 'failed'. Idempotent: same brief + same org reuses the same job_id, so retries don't fan out duplicate runs.

Dispatch to the QUANTITATIVE RESEARCHER — numerical analysis with full methodology context. Use for: briefs that turn on numbers done rigorously — "what is the documented effect size of X / what does the data say about Y / quantify the impact of Z". Every load-bearing number carries sample frame, sample size, measurement instrument, time window. Often answers with insufficient-evidence when underlying data is thin (negative findings are deliverable). Returns: 4-axis Quantitative summary (Value / Methodology rigor / Effect size / Robustness) + Numerical findings table + Methodology gaps + Sources. NOT for: topic landscapes (use dispatch_desk_researcher) / community language patterns (use dispatch_qualitative_researcher). ASYNC version: returns { job_id } immediately, the specialist runs durably on a Vercel Workflow (no 300s timeout). Use this version when the specialist is expected to take >90s. Call get_dispatch_result(job_id) periodically (respect wait_ms_hint in the response) until status === 'completed' or 'failed'. Idempotent: same brief + same org reuses the same job_id, so retries don't fan out duplicate runs.

Dispatch to the QUALITATIVE RESEARCHER — thematic synthesis from unstructured text (interviews, reviews, forum threads, customer language). Use for: "what are the 2-3 recurring themes in how D2C founders talk about X / what language is being used around Y / what are the patterns in customer reviews of Z". Every theme carries evidence count, triangulation status, ≥1 verbatim quote, outlier-check note. SOLVES the Reddit/X/Substack named-operator voice retrieval gap that legacy search tools could not fill. Returns: Corpus + Sampling + Coding methodology + 4-axis Themes table + Theme synthesis + Outlier voices + Saturation assessment + Sources. NOT for: quantitative effect sizes (use dispatch_quantitative_researcher) / multi-platform discourse mapping (use dispatch_social_listening_researcher). ASYNC version: returns { job_id } immediately, the specialist runs durably on a Vercel Workflow (no 300s timeout). Use this version when the specialist is expected to take >90s. Call get_dispatch_result(job_id) periodically (respect wait_ms_hint in the response) until status === 'completed' or 'failed'. Idempotent: same brief + same org reuses the same job_id, so retries don't fan out duplicate runs.

Dispatch to the SOCIAL LISTENING RESEARCHER — multi-platform community-signal interpretation. Use for: "what are practitioners saying about X across platforms / what jargon is emerging in field Y / what is the cross-platform discourse around brand/topic Z". Treats T3 community sources as primary data, distinguishes cross-platform patterns from single-platform noise. ≥3 platforms sampled per brief. Returns: Signal map (Signal / Platforms / Volume / Sentiment + recency) + Per-platform evidence trail + Cross-platform vs single-platform classification + Confidence flag + Sources. NOT for: single-source thematic work (use dispatch_qualitative_researcher) / numerical sentiment effect sizes (use dispatch_quantitative_researcher). ASYNC version: returns { job_id } immediately, the specialist runs durably on a Vercel Workflow (no 300s timeout). Use this version when the specialist is expected to take >90s. Call get_dispatch_result(job_id) periodically (respect wait_ms_hint in the response) until status === 'completed' or 'failed'. Idempotent: same brief + same org reuses the same job_id, so retries don't fan out duplicate runs.

Decode a specific video ad URL into its full structural formula — beat-by-beat breakdown, hook classification, behavioral psychology stack, creative format, runtime performance signals (active days on Meta Ad Library when available), and per-cut visual data. Takes one video URL plus an optional idempotency_key. Returns a job_id immediately; poll with get_decode every 15s until status is "completed" (typically 45-60s end-to-end). Use this when the user pastes an ad URL, names a specific competitor ad, asks "decode this" or "break down this ad" or "what makes this ad work", or wants sentence-level fidelity to one specific winner before writing a script with generate_adscript. Supports Facebook Ad Library, TikTok, Instagram Reels, YouTube Shorts, and direct .mp4 URLs. Costs 15 credits for videos ≤60s, 20 credits for 61-120s. Do NOT use to browse the corpus or find ads by category — use decoder_intelligence or adformula_intelligence (both free) for discovery. Do NOT use for image ads or static creative.

Retrieve the full decode bundle for a previously-submitted ad, or poll the status of a running decode job. Takes a single job_id (UUID returned by decode_ad). Returns either status="processing" (call again in 15s) or the completed payload — exact transcripts per beat, director's read, per-cut visual data (shot_breakdown), visual psychology, behaviour biases, beat structure, hook classification, and runtime fields (active days on Meta Ad Library when the source supports it). Use this immediately after decode_ad and every 15 seconds until the job completes. Also use this to re-fetch a decode any time you need the full bundle for script writing (Path B) or as the source_id for generate_adscript (source_type="decode"). Free — billing happens at decode_ad submit time, not on retrieval. Do NOT use to discover or list decodes — use decoder_intelligence for browsing. Do NOT use to start a new decode — call decode_ad first.

Build a complete creative intelligence profile of a brand from a single website URL. Takes a website URL (homepage, PDP, landing page) plus optional idempotency_key, force_refresh, and webhook_url. Returns a job_id immediately; poll with get_powersource every 3-5s (typically 60-90s total). The final payload contains 14 structured sections: identity, offer, selling_points, brand_story, brand_style, brand_assets, brand_voice, buyer_profile, 12 buyer tensions, marketing angles, emotional_arcs, ctas, proof_assets, and strategic narrative. Use this when the user says "analyse my brand", "load my brand", "build a strategy from my site", "what should my ads say", "decode this website", or pastes a homepage / competitor URL and wants a brand profile (not an ad decode). Also use this as the brand layer before calling generate_adscript — pass the returned powersource_id. Costs 100 credits. Re-scanning the same URL within your org returns the cached result free. Do NOT use for internal docs / PDFs / brand guidelines — use create_powersource_docs. For URL + docs combined (highest fidelity), use create_powersource_full. Do NOT use to decode a video ad — use decode_ad.

Retrieve the full creative intelligence profile for a previously-submitted PowerSource scan, or poll the status of a running scan. Takes a job_id (UUID returned by any create_powersource_* tool) plus an optional include_raw flag (admin-only). Returns either status="processing" with partial progress or the completed bundle: brand identity, offer, 12 selling points, brand voice rules, buyer profile, 12 buyer tensions, angles, emotional arcs, ctas, proof, narrative. Use this immediately after any create_powersource_* call and every 3-5 seconds until status is "completed". During synthesis, partial intelligence appears progressively (buyer archetype first, then tensions, then angles) — inspect each poll response, useful signal arrives early. Also use this to re-fetch a finished PowerSource any time you need the brand layer for downstream work. Free — billing happens at submit time. Do NOT use to start a new scan — call create_powersource_url, _docs, or _full first. Do NOT use to retrieve a video decode — use get_decode.

Build a complete creative intelligence profile from internal brand documents — creative briefs, brand guidelines, product specs, customer research, competitive analysis. Takes any mix of file_ids (from a previous upload), document_urls (public PDF/DOCX/TXT/MD links, up to 10), or documents_inline (base64-encoded files with filename), plus an optional context_url for layering live brand context (colors, fonts, current messaging) and optional idempotency_key. Returns a job_id; poll with get_powersource. Output shape is identical to create_powersource_url: identity, offer, selling points, voice, buyer profile, tensions, angles, emotional arcs, ctas, narrative. Use this when the user says "I have a brief", "here's my brand guidelines", "use this document", drops a PDF / DOCX / strategy deck, or when the truth lives in internal materials rather than the public website. The pipeline reads text only — convert PDFs to markdown before submitting via documents_inline when possible. Costs 100 credits. Do NOT use for URL-only scans — use create_powersource_url. For URL + docs combined (highest fidelity, triangulates public messaging against internal strategy), use create_powersource_full.

List images for a brand. Filter by PowerSource (this scan only, via powersource_id), by on-pack product_name (the vision tagger's read), by type (logo, product, product_cutout, hero, lifestyle, ingredient, packaging, certification, before_after, infographic, screenshot, video, general), or by is_primary_product. Use this BEFORE generating any image-based output so you pick from the brand's real assets, not generic stock. Returns asset_id, signed url, type, detected_product_name, is_primary_product, sources. Free, read-only. Paginated via cursor.

Upload an image to a brand by URL. The pipeline downloads it, runs the vision tagger (classifies type, detects product name, flags is_primary_product), stores it in the brand-assets bucket, and inserts a brand_assets row. Paid (vision tag credit). If vision tagging fails, the asset is still saved with type=general and can be retried via retag_brand_asset.

Delete one brand asset by asset_id. Removes the brand_assets row and (when the asset was uploaded rather than scanned) the storage object. Destructive — confirm with the user before calling. Use list_brand_assets first to find the asset_id.

Re-run the vision tagger on one brand asset. Reads the stored object when present (uploaded assets) or the original URL (scan-sourced assets), then updates type, detected_product_name, is_primary_product, description, and the other vision fields. Useful when the original tagger run missed or misclassified an image. Paid (vision tag credit).

List indexed brand documents for a brand. Each row carries the indexed signals (doc_type, summary, key_topics, classification_confidence, indexing_status) plus mime_type and size_bytes from the underlying file. Filter by doc_type (one of 19 values incl. voice_tone_doc, brand_guidelines, strategy_memo, customer_interview, pitch_deck, general_reference) or by indexing_status (pending, running, indexed, error). Use BEFORE read_brand_document to discover what context exists for a brand without paying the read cost. Free, read-only. Paginated via cursor.

Read one indexed brand document. Returns the indexed metadata (doc_type, summary, key_topics, entities, key_quotes) plus the document content, routed by mime: text/markdown, text/plain, text/csv are inlined as text; application/pdf and other text-shaped mimes return an Anthropic Files API file_id (attach via document source { type:"file", file_id } on the next turn); DOCX/PPTX/XLSX return a requires_code_execution marker (caller must enable code_execution_20260120 and attach via container_upload). Use AFTER list_brand_documents to pick the right document. Free, read-only.

List skills available in the Heista skill library. Returns name, description, domain (shared / image / video / research / strategy / copy / creative / generation), type (foundation / registers / models / methodologies), version, and source_folder (managed-agents / chat-agent). Returns frontmatter only — no body content (use load_skill for that). Filter by domain, type, or source_folder. Use BEFORE load_skill to discover what craft knowledge is available without paying the body-read cost. Free, read-only.

Load the full SKILL.md body for one skill by canonical dot-notation name (e.g. "research.foundation", "research.methodologies.desk-synthesis", "shared.registers.cinema-mode"). Returns frontmatter + body + content_hash. Verifies content_hash against the registry and surfaces drift if the registry is out of sync with disk. Use AFTER list_skills to pick the right skill. For register-type skills with references/ folders, follow with load_skill_reference to pull specific references. Free, read-only.

Load one reference file from a register-type skill's references/ folder (e.g. "m1-narrative.md" from "shared.registers.cinema-mode"). Only register-type skills have references/ — foundations and methodologies are inline content only. Find valid reference filenames in the parent SKILL.md's router table. Path-traversal protected. Free, read-only.

Web-grounded search via Perplexity Sonar Pro. Returns synthesized answer text plus a structured sources[] array (url + title) the caller can evaluate per the research.foundation four-tier source ladder. Optional recency_filter (hour/day/week/month/year) for fast-decay topics. Optional search_domain_filter (up to 10 domains) for triangulating against known-authoritative sources. Use this whenever a specialist needs current, web-grounded information — landscape scans, trend research, evidence queries, counter-evidence checks, named-entity lookups. Pair with the research.foundation skill (always-on craft baseline) and the research.methodologies.desk-synthesis skill (6-phase workflow) for production-grade output. The agent decomposes the brief into sub-questions BEFORE calling this — one focused query per call, not a multi-question batch. Cost is real (~$0.005-0.015 per query); the agent should budget calls per research.foundation §6 (fact-check 1-3, single comparison 3-8, landscape scan 8-20).

General-purpose web grounding via parallel.ai (Vercel AI Gateway). Returns synthesized text excerpts plus structured sources[] with direct URLs. Use for: topic landscapes, entity-deep teardowns, recency-sharp queries, named-vendor lookups, general fact retrieval. NOT for: Reddit/X/community discourse → use search_community. NOT for: numerical effect sizes or methodology-heavy fact-check → use search_research. The agent decomposes the brief into sub-questions BEFORE calling — one focused query per call. Optional after_date (ISO YYYY-MM-DD) for fast-decay topics. Optional max_results 1-20, default 10.

Community-discourse search via parallel.ai with optional platform filtering. Returns synthesized text excerpts plus direct URLs to real Reddit threads, X posts from named operators, Substack essays, LinkedIn posts, Facebook posts. Use for: "what are practitioners saying about X", recurring themes in founder voice, multi-platform discourse mapping, verbatim quotes from named individuals. Per Phase 3.5 empirical A/B (Docs/solutions/architecture-decisions/search-backend-architecture-jun04.md): this tool SOLVES the Reddit/X retrieval gap that perplexity_search fundamentally couldn't fill. Optional platforms[] to restrict (e.g. ["reddit","x","substack"]). Per social-listening-synthesis §3 sample ≥3 platforms per brief.

Structured fact-check + numerical research via Perplexity Sonar Reasoning Pro (Gateway-routed). Returns synthesized answer text plus structured sources[] with direct URLs to primary sources. Use for: specific numerical claims with methodology context, fact-check against primary sources, effect sizes + confidence intervals, earnings transcripts / SEC filings / research papers. Per Phase 3.5 empirical A/B: 2-3× cheaper than sonar-pro with comparable or better quality on structured research. Real Meta IR press releases + earnings transcripts on Desk. 17 cites on Quant. NOT for: Reddit/X/community → use search_community. NOT for: broad topic landscapes → use search.

Drill into a specific URL after search surfaces it. Returns the extracted text content plus metadata. Internal routing: PDFs hit Anthropic Files API for OCR + structured extraction; HTML pages are fetched + text-extracted via readability-style stripping. Use for: verifying a verbatim quote from a Reddit thread, reading a primary source in full (earnings transcript, research paper), drilling into a vendor product page after search surfaced the URL. NOT for: discovering new URLs — use search/search_community/search_research first. This tool takes a known URL only. Optional max_chars 100-50000, default 8000. SSRF-protected: private IPs + localhost blocked.

Dispatch to the DESK RESEARCHER — source-grounded synthesis on a topic landscape. Use for: "what is known about X / give me the landscape of Y / fact-check Z / synthesize the published evidence on W". Multi-source FACT/INFERENCE extraction with citation discipline. Vertical and geography agnostic. Returns: BRIEF restatement + NOT IN SCOPE + findings with FACT/INFERENCE/SPECULATION labels + [n] citations + Sources block. NOT for: trajectory questions (use dispatch_trend_researcher) / entity teardowns (use dispatch_market_analyst) / numerical effect sizes (use dispatch_quantitative_researcher) / community quotes (use dispatch_qualitative_researcher).

Dispatch to the TREND RESEARCHER — recency-dominant trajectory investigation. Use for: "is X a real trend / what is happening with X right now / where is X headed / what is driving X". Distinguishes trend from spike, signal from noise, real shift from echo chamber. Commits to falsifying conditions before searching. Returns: 4-axis Trend assessment (Reality / Magnitude / Direction / Horizon) + Current state + Baseline + trajectory + Drivers + Counter-signals + Sources. NOT for: static landscape questions (use dispatch_desk_researcher) / entity teardowns (use dispatch_market_analyst) / numerical analysis (use dispatch_quantitative_researcher).

Run a full research workflow via the Head of Research agent. The Head decomposes your brief into specialist sub-questions, dispatches the right combination of 6 specialists (desk, trend, market, quant, qual, social) in parallel via async dispatch, polls them to completion, judges output quality, and returns a structured synthesis. Use for: any source-grounded research request — fact-checking, vendor teardowns, trend assessment, quantitative effect-size analysis, qualitative theme extraction, cross-platform discourse mapping, or any combination. Wall time: 2-5 min typical. Returns: { synthesis, head_session_id, status, event_count, tool_uses, elapsed_ms }. NOT for: non-research requests (writing, coding, casual chat) — respond directly without calling this. Cost: $0.20-1.50 per call depending on brief complexity (specialist token spend + Anthropic session-runtime at $0.08/hr).

Dispatch to the MARKET ANALYST — entity-deep teardown of a named brand or vendor. Use for: "what is brand X / how does company Y work / decode competitor Z / teardown vendor W". Multi-axis extraction grounded in multi-class sourcing, plus defensible MOAT and credible GAP theses. Vertical and geography agnostic. Returns: 8-axis extraction (positioning / offer / audience / voice / pricing / distribution / proof / trajectory) + MOAT thesis + GAP thesis + Sources. NOT for: topic landscapes without a named entity (use dispatch_desk_researcher) / trajectory questions about a category (use dispatch_trend_researcher).

Dispatch to the QUALITATIVE RESEARCHER — thematic synthesis from unstructured text (interviews, reviews, forum threads, customer language). Use for: "what are the 2-3 recurring themes in how D2C founders talk about X / what language is being used around Y / what are the patterns in customer reviews of Z". Every theme carries evidence count, triangulation status, ≥1 verbatim quote, outlier-check note. SOLVES the Reddit/X/Substack named-operator voice retrieval gap that legacy search tools could not fill. Returns: Corpus + Sampling + Coding methodology + 4-axis Themes table + Theme synthesis + Outlier voices + Saturation assessment + Sources. NOT for: quantitative effect sizes (use dispatch_quantitative_researcher) / multi-platform discourse mapping (use dispatch_social_listening_researcher).

Dispatch to the DESK RESEARCHER — source-grounded synthesis on a topic landscape. Use for: "what is known about X / give me the landscape of Y / fact-check Z / synthesize the published evidence on W". Multi-source FACT/INFERENCE extraction with citation discipline. Vertical and geography agnostic. Returns: BRIEF restatement + NOT IN SCOPE + findings with FACT/INFERENCE/SPECULATION labels + [n] citations + Sources block. NOT for: trajectory questions (use dispatch_trend_researcher) / entity teardowns (use dispatch_market_analyst) / numerical effect sizes (use dispatch_quantitative_researcher) / community quotes (use dispatch_qualitative_researcher). ASYNC version: returns { job_id } immediately, the specialist runs durably on a Vercel Workflow (no 300s timeout). Use this version when the specialist is expected to take >90s. Call get_dispatch_result(job_id) periodically (respect wait_ms_hint in the response) until status === 'completed' or 'failed'. Idempotent: same brief + same org reuses the same job_id, so retries don't fan out duplicate runs.

Dispatch to the QUANTITATIVE RESEARCHER — numerical analysis with full methodology context. Use for: briefs that turn on numbers done rigorously — "what is the documented effect size of X / what does the data say about Y / quantify the impact of Z". Every load-bearing number carries sample frame, sample size, measurement instrument, time window. Often answers with insufficient-evidence when underlying data is thin (negative findings are deliverable). Returns: 4-axis Quantitative summary (Value / Methodology rigor / Effect size / Robustness) + Numerical findings table + Methodology gaps + Sources. NOT for: topic landscapes (use dispatch_desk_researcher) / community language patterns (use dispatch_qualitative_researcher).

Dispatch to the SOCIAL LISTENING RESEARCHER — multi-platform community-signal interpretation. Use for: "what are practitioners saying about X across platforms / what jargon is emerging in field Y / what is the cross-platform discourse around brand/topic Z". Treats T3 community sources as primary data, distinguishes cross-platform patterns from single-platform noise. ≥3 platforms sampled per brief. Returns: Signal map (Signal / Platforms / Volume / Sentiment + recency) + Per-platform evidence trail + Cross-platform vs single-platform classification + Confidence flag + Sources. NOT for: single-source thematic work (use dispatch_qualitative_researcher) / numerical sentiment effect sizes (use dispatch_quantitative_researcher).

Dispatch to the TREND RESEARCHER — recency-dominant trajectory investigation. Use for: "is X a real trend / what is happening with X right now / where is X headed / what is driving X". Distinguishes trend from spike, signal from noise, real shift from echo chamber. Commits to falsifying conditions before searching. Returns: 4-axis Trend assessment (Reality / Magnitude / Direction / Horizon) + Current state + Baseline + trajectory + Drivers + Counter-signals + Sources. NOT for: static landscape questions (use dispatch_desk_researcher) / entity teardowns (use dispatch_market_analyst) / numerical analysis (use dispatch_quantitative_researcher). ASYNC version: returns { job_id } immediately, the specialist runs durably on a Vercel Workflow (no 300s timeout). Use this version when the specialist is expected to take >90s. Call get_dispatch_result(job_id) periodically (respect wait_ms_hint in the response) until status === 'completed' or 'failed'. Idempotent: same brief + same org reuses the same job_id, so retries don't fan out duplicate runs.

Read a creative strategy in full by its powersource_id. Returns the same brand-merged bundle shape as get_powersource(data) — buyer profile, 12 behavioral tensions, angles, narrative direction, tone of voice, selling points, CTAs, proof, brand story, homepage data, offering — projected through the public PowerSource API serializer. Use this when you already have a powersource_id (from list_strategies) and want the full strategy payload in one call, without the job_id round-trip that get_powersource needs. Archived strategies are excluded by default (parity with list_strategies). Pass include_archived=true to read archived strategies. Read-only, free, account-scoped.

List tone profiles for a strategy. Today returns at most one entry — the tone_of_voice synthesized by the Tone of Voice Synthesis agent (POWER-mode bundles only). The shape is list-stable so future multi-tone bundles plug in without changing the contract. Use this to align generation with the brand-tied voice DNA before writing copy, hooks, or scripts.

Saved style configs picked into image-led Heists. Workspace = org-owned styles. Official = canonical Heista catalog (org_id IS NULL, is_canonical=true). Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one visual styles preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Heista-curated visual heists (style DNA + photography presets). Official-only today; workspace-saved presets are a future surface. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one visuals preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Structural references for script-led Heists. Workspace decodes (your video_sources scans joined with their video_scan_frameworks) + Heista-curated decoded ads from official_ad_heists. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one decoded ads preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Cluster-level structural formulas derived from decoded ads. Heista-curated; served as a generation parameter. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one ad formulas preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Static-ad references for image-led Heists. Workspace static scans + Heista-curated image ad heists. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one static ads preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Visual ideas you saved from prior generations. Workspace-only. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one saved visual ideas preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Reusable creative agents the Heist can pick as a handoff target — picked from the UI, callable as an MCP tool from Managed Agents. Workspace = private agents in the org. Official = public_template agents in any org. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one creative agents preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Seven-section creative-mechanism lenses the Creative Director chat picks at session start. The picked playbook substitutes Layers 3 + 4 of the system prompt — voice + foundation — for the session (the lens IS who the agent is). Workspace = private playbooks; official = the Heista-curated catalog. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one creative director playbooks preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Saved casting talent — a person you can re-use across Heists. The Models Heist saves them on click; future Heists can pick one as a brand-aware talent reference. Workspace = your saved castings. Official = Heista-curated drops across fashion, lifestyle, everyday, character, and creator buckets. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one models preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.

Cards the user bookmarked from Creative Director chat — directions, concepts, executions, brand platforms, art directions, visual sets. Surfaces in /library + the chat-side tray. Saves happen through the dedicated /api/creative-director/bookmarks route (NOT through /api/library), so is_savable is false here — the library surface is read-only. Read-only, free. Filter scope with only_workspace / only_official (mutually exclusive — same toggle as the in-app library lens). Page with limit + offset.

Get one saved cards preset by id, including its full body payload (framework, agent config, etc.). Call the matching list tool first to discover ids. Free, read-only.