Product updates
Changelog
New capabilities, improvements, and fixes to the CueFrame API — the video-composition layer your agents build on.
Budget ceilings, graphics mandates, and stage holds
New
- Budget ceilings that deliver —
budgetUsd on compose is now enforced end-to-end: a live spend_updated ledger streams as the job runs, and when the ceiling nears, the Director stops refining gracefully and ships the best version within budget instead of overrunning.
- Graphics mandates on beats — a brief beat can now carry
graphic: { role, items }. Mandated graphics are honored at planning time or the plan is refused with the exact fix — list items are carried verbatim into the composition.
- Stage-by-stage hold — compositions now pause at every stage (content, layout, motion) by default for your review; auto-progression is the explicit opt-in (
gates.autoApprove).
- Session-sighted previews —
preview-frame and score-composition accept a session_id, so in-session previews and scores always reflect the composition being edited, never a stale one.
- Motion presets — apply a motion preset to any approved card to instantly mint a tenant component with the right animation and alpha-layer form.
- Catalog-driven design briefs — designs draw from a deterministic catalog of candidates, so the same editorial role gets a consistent, on-brand selection every time.
- Direct upload — upload footage straight to storage with a presigned PUT and finalize it through a single API call.
Fixed
- Pinned caption emphasis is honored, not graded — if you pin
emphasisPhrases, the evaluator no longer penalizes your deliberate emphasis pattern.
- Composes fail fast and honestly — a candidate that cannot see or verify its own work fails loudly instead of silently retrying the whole pipeline from scratch; refinement rounds stop as soon as scores provably cannot improve.
- No more self-blocked jobs — a leaked session can no longer hold a lease that blocks its own job from running.
- Designed clips persist — card and component clips built during design now survive being saved and reloaded.
Compose fails loudly, and a batch of reliability fixes
Improved
- A stalled compose fails loudly — infrastructure failures and stalled plans now surface immediately instead of leaving a compose hanging until it times out.
- Earlier exits on low-quality runs — a compose run now terminates as soon as output quality drops below threshold, rather than burning its remaining budget on a run that cannot recover.
- Your content is never silently dropped — the planner can no longer remove content you supplied, and a compose that cannot see its own output is reported as a failed run instead of quietly producing something unreviewed.
Fixed
- Media processing retries no longer report a false error — retrying a media job that was already running returned a client error, which briefly marked the item as failed even though it was processing normally. It now reports the job as already running.
POST /v1/media/:id/process works again — the endpoint failed with an internal validation error instead of starting processing.- API key organization scope is verified server-side — an API key's organization is now confirmed against the creator's membership when the key is minted.
- Overlay geometry survives a save — an overlay clip's opacity, scale and rotation were dropped when a composition was saved and re-loaded. They now round-trip intact. (Overlay placement continues to use
region.)
- Reframe gaps caught at authoring time — a gapped reframe is now rejected immediately instead of surfacing a full turn later.
- Cinematic titles clear the letterbox — title text no longer clips into the letterbox bars.
Faster vision rendering and tighter error handling
New
- Leaner vision payloads — every render phase now sends a shrunk vision payload to the LLM, speeding up vision-based rendering.
- Fail-fast tool errors — infrastructure-class tool errors now fail a candidate immediately instead of quietly burning through the step budget.
Improved
- More resilient sessions — session idle timing is tuned so an active session is never reaped before its first heartbeat lands.
- Consistent authoring model — authoring and design agents now run on one explicitly pinned model, eliminating an unpinned fallback that could silently select a different one.
Fixed
- Alpha clips render correctly — designer-authored alpha clips are baked before workflow evaluation, so they show up correctly in the final render.
- Clearer errors on malformed clips — a malformed clip now surfaces its own error instead of being misrouted.
- Visible reframes on verify — crop trajectories are now derived for reactive render and verify, so reframes actually show up.
- Exclusions respected at plan time — graphics of an excluded kind are no longer planned in the first place.
- Compositionless projects can author graphics — projects with no starting composition now seed correctly so the designer can add graphics.
Briefs, checkpoints, and hosted review — the agent flow, rebuilt
New
- Briefs — persist your video's intent as a first-class object: goal, audience, format, duration, beats with frame-accurate moment pinning, captions with emphasis phrases, seeded graphics carried verbatim, hard exclusions, and gate configuration.
compose {briefId} honors every field or refuses loudly naming the one it can't.
- Checkpoints — a held compose now pauses on a durable approval packet: the stage artifact, storyboard stills, critic verdicts, and a cost quote before you approve. Answer with
resume — approve, revise with anchored comments (the run revises in place, with included revision cycles printed up front), or abandon. The full v1→vN history is queryable forever.
- Hosted review pages — set
gates.reviewMode: "hosted" on a Brief and every checkpoint mints a browser review link: a human approves or requests changes from the page itself — no account, no agent session — and the decision rides the same resume rail your agent uses.
- Score packets everywhere —
score_composition, Director gates, and consult now all mint the same addressable evidence object, so any verdict can be re-read later with get_checkpoint.
- Profile —
get_profile returns your org's taste in one read: brand kits, registered critics with their reference media, and the recent decision log. list_critics and delete_critic complete critic management.
- Reference-grounded critics —
register_critic accepts referenceMediaIds: the judge sees your own frames as the standard, making "consistent with our house style" a machine check in both the Director's gates and score_composition.
- Usage ledger —
get_usage lists every metered event with project and job attribution plus its list price, and job terminals stamp listUsd, so "what did this project cost this month" is a query.
- Clip suggestions become Briefs —
suggest_briefs promotes each finalized clip suggestion to a composable Brief with its source window pinned, ready for compose {briefId}.
Improved
- Sessionless evidence —
preview_frame, score_composition, and preview_component no longer need a session: the warm render box is a server detail that boots, is reused, and reaps itself, billed flat per operation. preview_component now returns a motion clip rendered with the same pipeline production uses.
- One terminal event per job family — webhooks converge on
media.completed, suggest.completed, compose.completed, and render.completed, each carrying its status (success or failure) in the payload, plus checkpoint.ready for held gates. Existing subscriptions were migrated automatically.
- A cleaner verb surface —
compose, new_composition, list_catalog, preview_frame, resume, and suggest_briefs replace their older names; retired names return a pointer to their replacement instead of an unknown-tool error. list_projects, delete_composition, and media library filters (kind, source, tag) round out steady-state workflows.
- Quotes before spend, on both paths — Briefs carry a journey estimate at creation, checkpoints carry the cost of the next step, and
validate_composition now returns the price to finish alongside its verdict.
- Faster verification — the composition judge runs on a faster model with cached rubrics, cutting scoring latency and cost without changing the criteria.
- Media context at a glance —
get_media_context now includes a presigned thumbnail beside faces and transcript.
A redesigned Guides library
Improved
- Redesigned Guides index — /guides now presents each guide as a cover card with category filters and a featured pick, instead of a plain list.
- Cleaner guide reading experience — every guide reads in a focused article layout with an author byline, a sticky on-page table of contents, and related guides at the end.
Catalog grows to 103 primitives with a browsable gallery
New
- Component gallery — browse the entire primitive catalog as real rendered examples at /components instead of static docs.
- 15 new launch-video primitives — the catalog grew from 88 to 103, adding dedicated AI-launch and launch-video templates plus a luminous flow-field silk background.
Improved
- Sharper visual taste pass — charts, kinetic typography, mesh backgrounds, and brush effects were rebuilt for cleaner, more polished output.
- Real transition wipes — seven component-wipe effects now run as proper transition presentations for smoother A-to-B cuts.
Fixed
- Uploads no longer cost credits — only exports and renders draw from your quota; importing footage is free.
- Cleaner error responses — requesting a malformed document id now returns a 404 instead of a 500.
- Primitive defects resolved — a batch of catalog issues found in audit, including preview and copy inaccuracies, has been fixed.
Self-serve site refresh, public status page, and clearer usage-based pricing
New
- Usage-based storage pricing — paid tiers now include storage overage billed at $0.03/GB-mo.
- Public status page — check platform health anytime from a status page linked in the site footer.
- Docs changelog timeline — browse the full release history directly on docs.cueframe.ai.
- Self-serve site refresh — the marketing site has a new self-serve-focused design.
Fixed
- Clearer render & generation pricing — included renders are now capped with pay-as-you-go overage, AI generation appears as its own line, and the tiers are easier to tell apart.
- Generation billed separately from rendering — AI generation now bills on its own line instead of counting against your included renders.
- FAQ matches the self-serve pivot — landing page FAQ copy now reflects the current self-serve plans.
- More reliable API-key checks — read-only API-key verification no longer contends with writes, avoiding request slowdowns.
Pay per tool call over hosted MCP
New
- Pay-per-tool-call over hosted MCP — connect an agent to CueFrame's hosted MCP endpoint and list the available tools for free, then pay per data-tool call with x402 — no API key required to connect. Payment rides with each call and settles automatically; a call made without payment returns a standard payment-required response so the agent knows exactly what to pay.
- Video-craft skill for MCP agents — agents driving CueFrame over MCP can pull a built-in composing-video skill that teaches how to produce a non-slop video in any format: ground the video in real footage, data, or UI; write kill-criteria before rendering; iterate on real rendered frames; and use composition verification as an independent second opinion before shipping.
Fixed
- Faster self-correction on broken components — when an authored component fails with a deterministic runtime error (for example a bad animation range), the preview now returns that error immediately in a readable form so the agent can fix and re-author right away, instead of silently retrying and stalling the edit.
Extract brand kits from a URL, search fonts, and edit in sessions
New
- Extract a brand kit from a URL — point at a website and pull its colors and branding into a reusable brand kit.
- Font search — search available fonts by name or intent and apply the result directly to your composition.
- Session-based editing — open an editing session, stage your changes in an isolated scratch, then commit to promote them or discard to throw them away — legacy projects included.
Improved
- Interactive playground drives real output — the font and caption-style selectors now produce real rendered output, and refreshed examples showcase text and custom graphics composited behind the subject.
Fixed
- Custom graphics render behind the subject — custom alpha-layer graphics are now first-class in the behind-subject depth order, so they sit correctly behind your speaker.
- Crop covers the silent lead-in — auto-reframe crop intents now start from the very beginning of a clip, covering the silent lead-in instead of leaving it unframed.
Get composition advice on demand
New
- Composition advice — a new advisory tool reviews your composition and returns grounded, actionable guidance before you spend a render.
Improved
- Steadier compose runs — compose no longer surfaces opaque internal errors and is less prone to intermittent flakes.
Fixed
- Reliable frame previews — still-frame capture is de-flaked with a render-time gate and automatic retry, so previews come back consistently.
Search stock media and keep on-screen text perfectly fitted
New
- Stock media search and import — search a catalog of stock video, images, and audio and import any result straight into your project, with attribution returned alongside each asset.
- Auto-fit text — every display-text primitive (titles, lower-thirds, subtitles) now measures, shrinks, wraps, and ellipsizes automatically, so copy always fits its frame without hand-tuning length caps.
- Runnable playground — a dedicated playground page gives you a live API console and a sandbox key to try calls in seconds.
- More in the free tier — the free tier now includes 10 GB of storage and brand kits.
Improved
- Sharper text behind your subject — text and captions placed behind a speaker now render with razor-sharp edges (down to individual strands of hair) and no color or quality degradation.
Fixed
- Billing on success, not on submit — composes, renders, and generations are metered when the work actually completes, not when you enqueue it.
- Clearer validation — invalid primitive parameters are caught at the write boundary with a typed, readable error instead of failing later during render.
Brand kits, multi-format derive, and a discoverable component catalog
New
- Brand kits — define your logo, intro/outro bookends, watermark, motion, spacing, and sizing once, and have them applied automatically at render time.
- Multi-format derive — reframe an existing composition into a new aspect ratio to produce a matching multi-format sibling.
- Discoverable component catalog — browse and place 87+ built-in primitives and scenes, filtered by tier, use case, and free-text intent, each with placement hints.
- Component workspace — eject a built-in primitive as owned source and scaffold, push, and sync your own custom components.
- Account and entitlements endpoint — see your plan, entitlements, and per-call cost ceilings up front, before hitting a payment wall.
- Browser sign-in over hosted MCP — connect an agent by logging in through a browser consent page, with no API key required.
Improved
- Smarter captions — pagination now respects sentence and segment boundaries, and captions clear during long pauses instead of lingering.
- Honest render lifecycle — renders report clear, terminal states and detect stalls, so you always know whether work finished or failed.
Fixed
- Custom graphics overlays — custom graphics now bake correctly at render time, and a fade-up animation no longer breaks with a zero distance.
- Caption backdrop — the legibility scrim is no longer drawn on videos that have no captions.
- Clearer import failures — failed imports now surface the underlying cause instead of an opaque error.
Pay-per-use billing and agent-native discovery
New
- Charge-on-delivery pricing — pay-per-use across composing, suggestions, media import, subject detection, and render, settled only when work is actually delivered.
- Agent-discoverable connection — a server card and manifest published at a well-known discovery path so agents can find and connect to CueFrame automatically.
- Standalone subject detection — trigger face and subject detection on its own, without running a full compose pass first.
Improved
- Consistent brand chrome — shared navigation, logo, and footer across the app surfaces.
Fixed
- Cross-tenant access closed — a session could be deleted across account boundaries; this is now blocked.
- Clip suggestions on silent media — media with no speech now returns a clear, typed error instead of an empty result.
Manage API keys and approve devices from the browser
New
- Web console — manage your API keys and account from a browser dashboard instead of the command line.
- Device sign-in — approve CLI and device logins through a dedicated device-authorization page (RFC 8628).
Improved
- CLI targets production by default — the CLI no longer locks to staging; commands run against your production account out of the box.
- Rebranded, tag-organized API reference — the reference is grouped by API tag so endpoints are easier to find.
Fixed
- API reference now resolves — the reference index and docs root no longer return a 404 or the scaffold home page.
CueFrame is live: the agent-native video API
New
- One API, driven by your agent — a curated tool surface over MCP (plus a CLI and SDK) lets an agent run the whole workflow: create a project, bring in media, compose, render, and get results back, with webhooks for every async event.
- Bring footage in, or generate it — import video, still images, or audio from a public URL, or generate media from a prompt (text-to-video, text-to-image, image-to-video, motion graphics, or auto), with content-addressed dedup so identical requests don't re-bill.
- Automatic transcription and perception — uploads are transcribed in 30+ languages with speaker diarization and per-word timing, and CueFrame detects faces and subjects so your captions, crops, and timing line up with what's actually on screen.
- Content-aware reframing — reframe any source to a new aspect ratio with a camera that smoothly follows the subject and zooms to fit, no manual keyframing.
- Captions that read like part of the shot — sentence-aware, per-word caption pages with emphasis styling, and the option to place captions and graphics behind the subject so they never cover a face.
- A built-in catalog to compose with — 87 primitives and ready-made scenes (lower-third, cinematic-title, color-grade, product-launch trailer, and more), filterable by use case, mood, and tier, plus the ability to author and preview your own custom components.
- On-brand by default — create or extract a brand kit (contrast-safe colors, neutral ramps, and brand fonts) and CueFrame applies your identity across every render.
- Render, review, and hand off — render to final video with cancel/retry and still-frame previews, or export an edit-ready timeline to Final Cut Pro (FCPXML) or Premiere.
- Pay-per-call, agent-native billing — per-plan entitlements and cost ceilings you can check before you spend, with pay-per-use settlement built into the API.