Clueso MCP Server

Add a screenshot or GIF to the article. Returns a screenshot_id to place in the article via update_article using `<SCREENSHOT screenshot_id>`. - kind="image": pass `timestamp` to capture a frame from the project's base video, OR pass `mcp_upload_id` (from the upload tool) to attach an uploaded image. One of the two is required. - kind="gif": pass `timestamp` (start) and `duration` (seconds). Captures a short GIF clip from the base video. Requires the project to have a video clip.

Get the article content as markdown. Screenshots appear as <SCREENSHOT {id}> placeholders. Returns the markdown text and a list of screenshot IDs.

Update article text content from markdown. Use <SCREENSHOT {id}> placeholders to position screenshots. Screenshots in the markdown are kept. Screenshots removed from markdown are deleted. Order in markdown defines final order. Get the current article first with get_article, modify the markdown, then call this.

Update an article image/GIF (screenshot), or inspect its current state. - Provide one or more of `zoom_factor`, `crop`, `shapes` to change metadata. - Pass `render=true` (with no other updates) to fetch the current details without mutating. To DELETE an article image/GIF: call `update_article` with markdown that omits the `<SCREENSHOT id>` reference. Unreferenced media is pruned server-side on save. There is no separate delete tool — the markdown IS the source of truth for what's present.

Create a new clueprint (reusable video template). A clueprint is a structured file directory with design notes, composition rules, and screenshots that can be used to generate videos. Use human-readable folder categories: design/ (colors, typography, backgrounds, animation), rules/ (composition, voiceover, editing), screenshots/ (rendered clip previews). See the "Creating Clueprints" section in the system instructions for the full structure. If source_guide_ids are provided, guide data is automatically extracted clip-by-clip into the clueprint as reference material.

Fetch clueprint data. Use `include` to control how much you pull back: - include="metadata": just the clueprint's name/description/tags/visibility/thumbnail. - include="tree": metadata + the full file tree (path, mime_type, and size per entry). Call this first when exploring an unfamiliar clueprint. - include="files": the contents of the files listed in `file_paths` (no metadata — call include='tree' first if you also need metadata). Text inline, binary as presigned GET URLs. - include="all": metadata + the file tree + the contents of every file in the clueprint (use sparingly on large clueprints). Calling this also logs the clueprint as "used" (fire-and-forget) so the workspace's recents list stays accurate — no follow-up use_clueprint call needed.

Update a clueprint's metadata and/or file contents in one call. Metadata fields (`name`, `description`, `tags`, `visibility`, `thumbnail_path`) are applied as a patch — only the fields you pass change. Pass at least one to update metadata. `files` is a list of file mutations: - Write text: `{ path, content }` - Write binary: `{ path, content, encoding: "base64" }` - Write from URL: `{ path, source_url }` (presigned URL, e.g. from get_clip with save=true) - Delete: `{ path, content: null }` You can mix writes and deletes in a single call. Existing files at the same path are overwritten.

Add many elements across one or more clips in a single tool call. Replaces the per-element tool — always batch. SEND LESS. Most batches repeat themselves — the same clip_index, element_type, font_setting, alignment or gradientSetting on every item. Hoist those into `defaults` and send them once; each item then carries only what differs, and overrides any key it sets (type_data merges one level). On a 20-30 element batch this typically cuts the call by a quarter to a half. `returning` shapes the response the way `select` shapes a get_clip read. defaults: { clip_index: 2, element_type: 'text', type_data: { font_setting: { font: 'Inter', weight: '600' }, alignment: 'center' } } elements: [ { name: 'headline', x: 960, y: 400, type_data: { text: 'Hello', fontSize: 120 } }, ... ] Use after calling get_element_schema to confirm the type_data shape per element_type. Items within one call are applied in order; returns one result entry per input item so partial success is fine. Reuse instead of rebuilding: an item may pass `component_id` INSTEAD of element_type/type_data to insert a saved component from this workspace exactly as stored — no generation, instant — returning every new element_id plus its parameter_schema; set its content afterwards with update_elements(type_data.parameter_values). Find them with find(type='element_components'). Use `base_component_id` (with element_type='animation' and a prompt) only when you want a NEW variant rather than that component. Z-order: list position IS z-index — later renders on top — and a new element goes on top by default. Pass `insert_at: 'back'` to put it behind what's already there, or `insert_at: {before: id}` / `{after: id}` to land next to a specific element — that also puts the new element in THAT node's group, which is the only way to place inside one. `reorder_elements` rewrites the whole order in one call, grouped clips included. `group` + `insert_at: {before|after}` together: what the reference is relative to decides what gets positioned. Point at a node INSIDE the group and the ELEMENT takes that slot among its new siblings. Point at one OUTSIDE it and the GROUP takes that slot, with the element inside — which is how you put a backdrop pattern of N elements at a chosen depth as one hideable unit: hoist both into `defaults` and the first item seats the group, the rest just join it. The one refusal left is naming a reference outside a group that already exists somewhere else, since moving an established group is not what adding one element to it should do; seat the group where you want it in the call that CREATES it, because there is no reposition afterwards. `front`/`back` with a `group` follow the same principle: they address the GROUP only while the element is alone in it — the call that creates it — and once the group has other members they address the ELEMENT within the group's interior, so they will not move an established group either. Grouping: pass `group: "<name>"` to keep a unit together (a card and its label, a stat and its caption) so the user can move or hide it as one thing. Items in one call sharing a name land in the same group, and a later call with that name adds to it. Grouping never changes coordinates. It does affect z-order: a group's members render contiguously at the group's slot, and a NEW group takes the slot of its first member, so grouping already-adjacent elements keeps their z-position while grouping scattered ones pulls them together at the lowest member's slot. Concurrency: within ONE call every element lands in a single save. Across calls the conflict domain is the CLIP's element list, not the individual element, so you can fan this tool out across parallel subagents targeting DIFFERENT clips. Two concurrent add_elements calls on the SAME clip are NOT safe: every add claims that clip's element list, so the later call is REJECTED ('changed since this edit was based on') and nothing it sent is written — batch all of a clip's elements into ONE call instead. A rejection is not last-write-wins: re-read and re-apply. Do NOT run it concurrently with a whole-clip or whole-project mutation on the same guide (update_clips on that clip, add_clips/remove_clip/split_clip/duplicate_clip, add_audio, update_project) — those rewrite a larger scope and would clobber the element. Element-type quirks (handled per-item): • zoom → x/y/width/height are ignored; use centerX/centerY in type_data • image → provide x/y/width/height (the clip is located by clip_index; the clip_id input is accepted but unused) • animation → x/y/width/height default to the full canvas if omitted (the clip is located by clip_index) For everything else, x/y/width/height are required. Animation: pass a top-level `keyframes` array (sibling of x/y/type_data, NOT inside type_data) — entries are { timestamp, positionX?, positionY?, width?, height?, interpolation? } in canvas pixels. positionX/Y use the SAME alignment-aware origin as the element's x/y. Text caveat: width/height are not keyframable on text — animate its size with fontSize (letterSpacing/lineHeight/padding* are also keyframable). Position origin: for TEXT `y` is the vertical CENTRE when centre-aligned and the TOP otherwise; every other type uses the top; `x` is the left edge, except TEXT where alignment picks it: left→left, center→centre, right→right. To centre a label in a row, pill, chip, card or beside an icon: pass the container's centre line as y with y_anchor:'center'. The server centres the measured text in it, so it is one number and it holds for wrapped text. Add type_data.fit{max_height} whenever the text could wrap, so the box stays inside the space you gave it. y_anchor is a directive for computing the stored y, not a property that sticks — repeat it on any later call that moves the same label.

Remove many elements across one or more clips in a single tool call. One entry per element ({clip_index, element_id}). Concurrency: parallel-safe (conflict domain: the individual element) — same as update_elements. Each removal is a granular element_remove patch scoped to that element id, and the whole batch lands in ONE save. Fan out across subagents as long as they touch DIFFERENT element ids. Two concurrent edits to the SAME element id do not last-write-win — the later one is REJECTED and nothing is written; re-read and re-apply. Do NOT run concurrently with whole-clip/whole-project mutations on the same guide (update_clips on that clip, structural clip ops, add_audio, update_project). To remove an audio track (not an element), use remove_from_project(target='audio').

Update many elements across one or more clips in a single tool call. Replaces the per-element tool — always batch. SEND LESS, GET BACK LESS. `defaults` carries what every item shares (clip_index, element_type, and type_data entries like font_setting or alignment) so you send it once instead of per item; each item overrides any key it sets. `returning` declares what comes back, the way `select` shapes a get_clip read — omit it for a sensible default, or pass [] for just the counts. defaults: { clip_index: 1, element_type: 'text', type_data: { font_setting: {...} } } returning: [] → { total, succeeded, failed } only returning: ['layout'] → + the measured text fit per element returning: ['qa.codes'] → + the clip audit with prose stripped from non-errors (any returning value → error findings regardless; only warnings are opt-in) Each item can update position/timing fields alone, or type-specific fields via type_data (use get_element_schema in 'update' mode to see what's settable for an element_type). Items within one call are applied in order; partial success is fine. You do NOT need clip_index here — an element_id is unique within the project, so the server locates the clip itself. Pass it only as a hint; a wrong one is corrected rather than rejected. Grouping: pass `group: "<name>"` to move an element into a named group (created on demand), or `group: ""` to pull it back out to the clip root. Several items sharing a name collect into one group — the way to tidy loose elements into units the user can move or hide together. Grouping never changes coordinates. It does affect z-order: a group's members render contiguously at the group's slot, and a NEW group takes the slot of its first member, so grouping already-adjacent elements keeps their z-position while grouping scattered ones pulls them together at the lowest member's slot. Customising a component: after add_elements(component_id=...), set its content here with type_data.parameter_values on the animation element it returned. Concurrency: parallel-safe (conflict domain: the individual element), same as add_elements — fan out across subagents as long as they touch DIFFERENT element ids — those carry different entity paths and both land. Two concurrent edits to the SAME element id do NOT merge and do NOT last-write-win: the later one is REJECTED ('changed since this edit was based on'), nothing is written, and you must re-read and re-apply. Do NOT run concurrently with whole-clip/whole-project mutations on the same guide (update_clips on that clip, structural clip ops, add_audio, update_project). Keyframes: pass a top-level `keyframes` array (sibling of x/y/type_data, NOT inside type_data) to set, or null to clear. Per-entry shape: { timestamp, positionX?, positionY?, width?, height?, interpolation? } in canvas pixels; positionX/Y use the element's alignment-aware origin. Text elements: pass `fit_to_lines: N` to run an automatic widen + font-shrink pass after the entry's regular update lands. The server reshapes the element so the rendered text wraps to at most N lines without overlapping its neighbours.

Create a new blank Clueso project. Returns the projectId and initialClipId — the project starts with one blank clip already at index 0 (default short duration). Use update_clips to resize/retitle clip 0, add_clips(kind='blank') to append more clips, add_elements to place content, and add_audio to attach a project-level music track. Optionally place the project inside a folder (use find(type='folders') to find folder IDs) and set a description at creation time.

Get a summary of a project: canvas dimensions, per-clip summary (ids, indices, durations, element counts, voiceover_volume), and `audio_tracks` — every music/SFX entry on the project flattened to the same wire shape `add_audio` accepts (id, src, name, source_duration, guide_start_time/end_time, music_start_time/end_time, volume, loop, fade_in, fade_out). To copy an audio track to another project: read it from `audio_tracks` here and feed the same fields to `add_audio` on the target project. Call this to understand what's in the project before making changes.

Update a project's metadata: rename, change description, move to a different folder, or change aspect ratio. Only fields you provide are changed. To move a project to the workspace root, pass folder_id="" (empty string). Aspect ratio: pass any "W:H" (positive integers). The canvas is fit inside 1920×1080 keeping the ratio, so element pixel coordinates use the resulting canvas. Common values: - "16:9" → 1920×1080 (landscape, YouTube/web — default) - "9:16" → 608×1080 (portrait — TikTok/Reels/Shorts) - "1:1" → 1080×1080 (square — Instagram feed) - "3:4" → 810×1080 (portrait card) - "4:5" → 864×1080 (portrait — Instagram feed) Element coordinates are stored as fractions of the canvas, so existing elements reflow to the new canvas automatically — no element coordinates are rewritten. Concurrency: whole-project mutation (conflict domain: the entire project) — serialize. Do not run it in parallel with ANY other mutation on the same project_id, including element/voiceover edits; run them one at a time. (Mutations to different projects run in parallel freely.)

Add a music track or sound effect that plays as a project-level audio layer over all clips. Does NOT consume a clip slot. Use this for background music or a soundtrack that should play under the existing timeline. (For inserting an audio file as its own sequential clip with auto-transcription, use add_clips(kind='video') with an audio mime_type instead.) - type="music": requires src, name, source_duration. `src` is either the `src` from a find(type='music') result, or an mcp_upload_id (prefixed `mup_`) from upload_file — upload ids resolve to a presigned URL server-side. For a Freesound result the `src` is an opaque handle; pass it through unchanged and Clueso downloads + hosts the original server-side (do not try to fetch it yourself). Get `source_duration` from the find result, or from analyze_audio mode='features' (data.duration_ms / 1000). Optional: guide_start_time, guide_end_time, music_start_time, music_end_time (trim the source track), volume, loop, fade_in, fade_out. - type="sfx": pass the `src`, `name`, and `source_duration` from a find(type='sfx') result, plus guide_start_time (Freesound `src` is an opaque handle, resolved server-side just like music). Optional: volume. (Legacy: a fixed-library `sfx_key` like 'whoosh' still works instead of src.) Concurrency: whole-project mutation (conflict domain: the entire project) — serialize; do not run in parallel with ANY other mutation on the same project_id (including element/voiceover edits). Size your clips BEFORE calling add_audio, or pass guide_end_time explicitly — the default ('end of project') is captured at call time and will NOT auto-extend if clips are added or resized afterwards.

Mutate fields on an existing music/SFX track by id. Read existing tracks via `get_project.audio_tracks` — the per-entry shape there matches this tool's args, so you can flip a knob (volume, fades, timing, loop) and write it back without losing the track's id. Only fields you pass are changed; omitted fields are preserved. To delete a track entirely, use `remove_from_project(target='audio')` instead.

Read one clip: its elements (positions/sizes in canvas pixels), voiceover (text, voice, duration, voiceover_volume), background and transition. Pass `render` to also get a PNG of the frame. ASK FOR WHAT YOU NEED. A full read is large — on a dense clip the per-word voiceover array and the element type_data blobs dominate it, and repeated full reads are the main way a long session runs out of context. `select` returns exactly the parts you name: select: ['elements.x','elements.y','elements.width','elements.height'] → geometry only, to fix a layout select: ['elements.name','elements.start_time','elements.end_time'] → a timing pass select: ['voiceover_words'] → word timings only, to sync visuals to narration select: ['elements.textdata','voiceover_words'] → rewrite copy against the VO select: ['elements'] → whole element rows, no words select: ['groups'] → group rows only, to get a group_id for update_groups select: [] → no JSON at all (pair with render for the PNG alone — smallest read) (omit select) → everything; fine for a first look, expensive to repeat `render` is the other output, and it is separate from `select`: `select` shapes the JSON, `render` produces a PNG. render: {} → the frame at t=0 render: { timestamps: 2.5 } → the frame 2.5s into the clip render: { timestamps: [0.5, 2, 4] } → those three moments as ONE labelled grid render: { timestamps: [...], layout:'separate'} → the same moments as full-size images (~4x the tokens) render: { save: true } → also uploads the frame and returns presigned_url render: { max_width: 1280 } → a sharper frame when you must read small print select: [], render: {} → the PNG alone, no JSON select: ['elements'], render: {} → element rows AND the frame A single frame renders 960px wide by default — legible for this design system and about half the tokens of a 1280 frame. A GRID defaults to 1280, because that is the width of the whole grid and a third of 960 would leave each cell unreadable. Either way `max_width` caps the image you get back; raise it only to read genuinely small print. Omitting `render` renders nothing. `timestamps`, `layout` and `save` live inside it because they only mean anything for a render — there is no way to ask for them without asking for the image. `element_ids` is the other axis: it picks WHICH element rows come back, independently of `select`. Combine them for the leanest read — e.g. element_ids: ['el_9'], select: ['elements.x','elements.y']. Element shape: universal wrapper fields (id, element_type, name, x, y, width, height, start_time, end_time, rotation) plus type-specific data (textdata/shapedata/imagedata/videodata/zoomdata) plus an optional `keyframes` array when animated. Keyframes come back in the same flat wire shape add_elements takes — { timestamp, positionX?, positionY?, width?, height?, interpolation? } in canvas pixels — so you can round-trip read → edit → update_elements without reshaping. Clip-level fields include `transition` (the current transition object — sibling of the update_clips `transition` arg; null if none) and `voiceover_words` (per-word timestamps, `{word, start, end, punctuated_word}` with start/end in SECONDS; null on clips with no transcription).

Remove a clip from the project by index. Cannot remove the last remaining clip. After removal, all clips with index > clip_index shift down by 1. When removing multiple clips by index, remove from highest index to lowest (or capture clip_ids first and re-derive indices between calls). Concurrency: whole-project mutation (conflict domain: the entire project) — serialize. Do not run it in parallel with ANY other mutation on the same project_id, including element/voiceover edits; run them one at a time. (Mutations to different projects run in parallel freely.)

Create new clips on a project. One tool, three sources: - kind="blank": empty slide clips you'll populate with add_elements / generate_media. Pass `clips: [{title?, duration?}, ...]`. Note: create_project already creates clip 0; calling with N blank entries gives N+1 clips total. - kind="pptx": extract slides from an uploaded .ppt/.pptx. Pass `mcp_upload_id` (from upload_file), optional `slide_indices`. Each slide becomes a slide_clip with native text/image/shape elements pre-populated. - kind="video": cut clips from an uploaded video/audio. Pass `mcp_upload_id` and optionally `mime_type` (default video/mp4), `title`, `voiceover_text`, `cuts: [{start_seconds, end_seconds, heading?}]`, `auto_elements`. Without `cuts` the whole upload becomes one clip. Pipeline (download → transcribe → insert) runs in the background; subsequent calls on the same upload hit a 24h cache and return synchronously. (To overlay a video ONTO an existing clip — picture-in-picture, b-roll, background loop — rather than insert it as its own clip, use add_elements(element_type='video') instead.) All variants accept `after_clip_id` to control insertion position; appends at end if omitted. Concurrency: whole-project mutation (conflict domain: the entire project) — serialize. Do not run it in parallel with ANY other mutation on the same project_id, including element/voiceover edits; run them one at a time. (Mutations to different projects run in parallel freely.)

Update one or more clips in a single call — background, transition, duration, title, visibility, position, and (video clips) the footage transform via video_styling. Always batch. Each entry: { clip_index, duration?, title?, background_color? | background?, transition?, hidden?, position?, voiceover_volume?, video_styling? } - voiceover_volume: 0-100. This is how you make a clip SILENT — set 0 to mute the narration while keeping the text and its word timings. get_clip and get_project report the current value, so read it before muting to restore it later. There is no "delete voiceover" tool; muting is the supported way. - background_color: hex (keeps the existing pattern/image, changes color). background: full object from find(type='backgrounds') — replaces the whole background. - transition: { type, duration, direction?, color? }, or null to remove. Set on the OUTGOING clip — the transition plays at the END of clip_index, into the next clip. Types: circleWipe, rainbowWipe, cornerWipe, stack, colorWipe, dissolve, pushPull, fadeToColor. Directions: rainbowWipe/cornerWipe take topLeft|topRight|bottomLeft|bottomRight; colorWipe left|right; stack/pushPull left|right|top|bottom. - position: move the clip to a new zero-based index. Moves apply FIRST (in array order), then property patches — keep reorders in a separate call from property edits so indices stay predictable. - video_styling (VIDEO CLIPS ONLY — ignored on slide clips): reposition/resize the clip's recording on the canvas, e.g. into a designed window slot. { position?: [cx, cy], size?: [w, h] } as canvas fractions 0–1 ([1,1] = full-bleed, position is the footage CENTER), plus crop? and border?. Fields merge onto the clip's existing video_styling. Concurrency: parallel-safe per clip — all property updates land in ONE save. Safe to fan out across DIFFERENT clips. Do NOT run concurrently with element/voiceover edits on the SAME clip, or whole-project mutations (add_clips, add_audio, update_project). When adding audio + resizing clips, run update_clips BEFORE add_audio.

Clone one clip — within the same project, or **from another project** — into a target project. - Same project: pass `target_project_id` and `source_clip_id` (omit `source_project_id`). - Cross-project: pass `target_project_id`, `source_project_id`, `source_clip_id`. The source clip's S3 assets (voiceover audio, original video footage, generated video, etc.) are re-hosted into the target guide's S3 namespace, so the new clip is independent of the source — deleting the source project later won't break it. Insertion: pass `after_clip_id` to place immediately after a specific clip in the target. Omit to append at end. Returns the new clip_id and its final index. Concurrency: whole-project mutation (conflict domain: the entire target project) — serialize; do not run in parallel with any other mutation on the same target_project_id.

Clone an entire project — every clip, voiceover, asset, and metadata field — into a fresh project. The copy has its own S3 namespace, so changes or deletions to the source never affect the duplicate. Defaults: - new_title omitted → "{source title} (copy)" when no folder_id is given (or it's empty) and the source stays in the same org; pass an explicit non-empty folder_id and the title is preserved. - folder_id omitted → copy lives in the same folder as the source. Returns the new project's id and title. Concurrency: whole-project read+create — serialize against other mutations on the same source_project_id; do not run them in parallel.

Fetch the outcome of a run_script that did not finish inline. Waits for it rather than returning immediately, so one call is usually enough — set `wait_seconds` to how long you are willing to hold. Returns { job_id, status, result?, error? }. `status` is processing, completed or failed; `result` carries the same { ok, result, stdout, ops } the inline reply would have. The script runs to completion whether or not you poll.

Write a JavaScript program that calls this server's other tools. You have the whole language: loops, arithmetic, functions, conditionals, and values carried from one call into the next — all running next to the tools instead of across the conversation. Use it whenever code says the thing more directly than a sequence of separate calls would, which is often. COMPUTE, don't hand-write. Anything you would otherwise work out in your head and type as literals is better computed here — eased keyframe tracks, staggered start times, grid coordinates, derived palettes, positions from measured text widths. This is usually what makes motion look right: sample a curve at ten points and emit the values, rather than guessing four. DON'T COMMENT THE SCRIPT. Nobody reads it — it runs once and is gone. Comments, blank lines and explanatory names are pure cost here. Write it dense. The one thing to remember: inside a script you see only what you `return` or log, so a tool's own rich output (layout measurements, layout_qa, new ids) has to be surfaced deliberately. That also makes a single-call script worthwhile when a read is fat — `find` takes no projection argument, so logging the 200 characters you need is a real saving. A fat read is usually better narrowed at the source than filtered here: get_clip takes `select`, and get_element_schema takes `fields` (name what you're setting and a 15KB schema becomes ~200 bytes). get_element_schema returns TEXT by default, which you cannot index into — pass `format: 'json'` when you want to compute over the schema rather than log it. CALLING TOOLS Every tool except run_script and get_script_job is a global function taking exactly the arguments it takes normally, and returning its parsed result (those two are excluded so a script cannot recurse into itself or poll its own job). Calls are synchronous — do NOT use async/await, and there are no imports. `sleep(ms)` waits, synchronously like everything else here. Use it to poll a generation: `generate_media` and `voiceover_batch` return before their work lands, and `get_clip(select:['busy'])` says what is still being written. It is clamped to the script's remaining time, so a long sleep ends the run rather than overrunning it — and it holds one of the few concurrent script slots while it waits, so poll on the order of seconds, not milliseconds. const p = create_project({ title: "Launch" }); const made = add_clips({ project_id: p.projectId, kind: "blank", clips: [{ duration: 4 }] }); Use the BARE tool name. If your client shows these tools under a prefix, the prefixed form works too — mcp__clueso_connect__add_elements and clueso__add_elements both resolve to add_elements. `tools()` lists every callable name; `call(name, args)` invokes one by a name computed at runtime. WHY IT IS CHEAPER What you construct never passes through the conversation. Build an array in a loop and send it in ONE batched call — a 24x24 dot grid is six lines here versus 576 elements of JSON: const els = []; for (let r = 0; r < 24; r++) for (let c = 0; c < 24; c++) els.push({ x: c * 60, y: r * 60, width: 8, height: 8, type_data: { backgroundColor: "#C462F5" } }); add_elements({ project_id, defaults: { clip_index: 0, element_type: "rectangle" }, elements: els }); READ, THEN WRITE — this is how you edit in bulk: const clip = get_clip({ project_id, clip_index: 0, select: ["elements.name", "elements.y"] }); const captions = clip.elements.filter(e => e.name.startsWith("caption")); update_elements({ project_id, defaults: { clip_index: 0 }, updates: captions.map(e => ({ element_id: e.id, y: e.y + 40 })) }); WHAT COMES BACK `return` a value to hand it back, and `console.log` anything you want to see — that output is all you pay for, so log summaries, not payloads. `print_json(obj, maxBytes)` logs an object and truncates it for you, which is safer than hand-rolling JSON.stringify(...).slice(...) at every call site. You get `{ ok, calls, writes_applied, result, stdout }`. `writes_applied` matters because a script is NOT a transaction: if it throws on call nine, the first eight writes already landed, and that count is how you tell. A FAILED TOOL CALL THROWS. It does not return an error object, so a failure stops the script instead of letting it run on against bad state. Wrap a call in try/catch only when you genuinely intend to continue. You get every tool's normal result, so `succeeded`/`failed` counts, per-element `layout` (font_size_px, text_width_px, natural_width_px, line_count, fits_width/fits_height/fits plus `adjusted` when your height was replaced — widen to natural_width_px, not text_width_px) and `layout_qa` findings are all readable in the script. Check them and react. NO IMAGES are returned inside a script — a render in a loop would flood the reply. But you are not blind: `get_clip({..., render: { save: true }})` gives the script a `presigned_url` and `s3_key` for the rendered frame, so it can render many frames, keep the URLs, and hand them on (update_clueprint takes one as source_url). `_images_omitted` counts the frames withheld; when the reply would be a bare array it arrives as `{items, _images_omitted}`. Call get_clip yourself, outside a script, when YOU need to look at one. FINISHING Fast scripts return their result here. If one is still running after `wait_seconds` (default 45s) you get `{ job_id, status: "processing" }` and it keeps running — poll `get_script_job`. Set `idempotency_key` on anything that builds, so a retry cannot run it twice — a key that already ran is never re-executed, failed runs included. LIMITS: 200 tool calls, 45s of run time, 16KB each of logged output and returned value. `dry_run` runs reads for real and only records writes. Anything costing credits errors as it normally would.

Analyze an audio file. Modes: - transcript: Speech-to-text with word-level timestamps — returns `{language, full_text, words}` where each word is `{word, start_ms, end_ms, punctuated_word}`. ⚠ MILLISECONDS here; get_clip's `voiceover_words` gives the same words in SECONDS as `{word, start, end}`. Use for caption timing, script editing, or word-level visual sync. Works on narration and on sung lyrics: the provider is picked from the content, and if it returns nothing the other one is tried automatically. So an empty `words` means the audio genuinely has no intelligible speech, not that the wrong engine ran. - silences: returns `{silences, total_silence_ms}`, each span `{start_ms, end_ms, peak_db}`. Use for auto-trim, smart-split, or pause-aware editing. - beats: returns `{tempo_bpm, beats_ms}` — beats_ms is an array of MILLISECOND positions. Empty for non-music. Use to align animations/zooms to a beat. - features: returns `{duration_ms, peak_db, rms_db, lufs, has_music, has_speech}` — loudness plus two INDEPENDENT hints — `has_music` (beats hold a steady period) and `has_speech` (there is audible content above a loudness floor; it is not a speech detector). Both can be true at once, e.g. narration over a music bed. Cheap dispatcher — call once to pick a mode, or for duration_ms to pass as add_audio's source_duration. When the words matter, trust transcript rather than these flags. Source: provide exactly one of mcp_upload_id (from upload tool) or file_url (already-public URL, e.g. a find(type='music') result or Clueso CDN file). Optional time range: range_start_ms / range_end_ms crops the analysis window.

Run the agentic auto-sync pipeline against a clip with a source recording. Detects natural sync points (UI state changes, narrated steps) and inserts sync-marker nodes into the clip's transcript. Async: returns immediately with a status enum from the pre-flight; sync-marker nodes appear in the transcript a few seconds later. Poll get_clip if you need to verify. Capacity: capped at 3 concurrent runs platform-wide. Returning status='success' means the job was accepted, not that it finished. Sync points are required input for voiceover TTS on video clips — without them, the TTS has no per-step pacing reference. (See resource clueso://docs/sync-points for the full model.)

Post a review comment on a project, attributed to "Clueso AI". Use this to leave feedback, suggestions, or notes — either at the project level (no clip_id) or pinned to a specific clip with an optional timestamp inside that clip. Pass `parent_comment_id` to REPLY to an existing comment rather than open a new thread. Read what is already there with `get_comments` first: answering a reviewer in their own thread is what makes the note resolvable, where a fresh top-level comment just adds noise.

Read the review comments on a project — the other half of `add_comment`, which could post but never read. Comments are how a human reviewer tells you what is wrong with a video: a note pinned to a clip and a moment inside it. Read them before an editing pass so you act on what was actually asked for, and read them again after a build if a reviewer has seen it. Threads come back nested: each top-level comment carries its `replies`. Reply into a thread with add_comment(parent_comment_id=...) rather than posting a new top-level note — a reviewer reading their own thread is how they see that their point was handled. Resolved threads are hidden by default, so a plain read is "what still needs attention".

Clueso's house style for composing video clips — how to think about concept, typography, color, composition, motion, and pacing. Read it before composing or editing a clip's visuals when you're NOT following a clueprint, or to fill gaps a clueprint leaves silent. General craft, not clueprint-specific.

Return the schema for an element type's type_data, as TypeScript declarations. Call this before add_elements / update_elements when you don't already know the field shape for the element_type you're placing. Pass element_type='group' for the GROUP surface instead — what a group is, how its pivot works, and the nine properties it can animate.

Estimate voiceover duration for one or many clips at once (~150 words/minute). Pass an array of voiceover texts — one per clip. Returns durations in the same order, ready for add_clips.

Kick off a video export of the project. Returns immediately with an export_id; the export pipeline runs in the background. The user can find the rendered video in the Exports tab in the editor. Defaults: 1080p / 30fps / no captions / English. Pass overrides only when needed.

Find any resource in Clueso by type, optionally filtered by name or exact id. One tool for listing and searching across the workspace. type: • projects | folders | clueprints | workspaces • backgrounds | voices | image_gen_style_packs | element_components (voices carry `gender`, `tags`, `accent`, `preview_url` and — where one has been written — a free-text `description` of tone and pace, which is the field actually worth choosing on. `tags` is a small closed vocabulary, so it groups voices rather than telling them apart. Choose deliberately: voiceover is the film's clock, and regenerating speech later rescales the clip and every element time and keyframe with it.) • images | videos | music | sfx — media; each result carries a `source` ('org' = your saved-media library, 'stock' = a stock/curated provider). Scope with `source`, pick the library with `provider` (see below). Stock results are a short described shortlist — pick the best fit and use its `src`. Stock video results also carry `safe_src` and a `video_files` tier list with one entry marked `recommended` — use `safe_src` (or the recommended tier) in add_elements; tiers above 1080p can exceed its ~200MB source cap and fail. For a Freesound music/sfx result, `src` is an OPAQUE handle (not a playable URL) — pass it straight to add_audio and the original is fetched + hosted by Clueso server-side; a `preview_url` is included only so you can tell what it sounds like. (image_gen_style_packs = generation style presets for generate_media kind='image' style_id; element_components = saved components (e.g. animations) from THIS WORKSPACE only — there is no community library for components (unlike clueprints); each reports param_keys. Insert one AS-IS with add_elements(component_id=...), or generate a variant from it with base_component_id.) Filters (all optional): • query — for stock media it's the search phrase (real semantic search for provider='clueso'; provider keyword search otherwise). For clueprints a query runs a relevance-ranked search across your workspace + the global community library (search_summary, relevance_reason, tags, is_community, fork_count). For everything else it's a case-insensitive name substring. • provider — which stock library to search (ONE call, no merging). Choose by strength: images → 'pexels' (default; realistic photography) or 'pixabay' (illustrations, vectors, icons, clip-art — set image_type) videos → 'pexels' (default; real-world footage) or 'pixabay' (motion graphics — set video_type='animation') music → 'clueso' (default; our curated, brand-safe library with the best descriptions/search — try this FIRST) or 'freesound' (niche/genre tracks) sfx → 'freesound' (default; vast sound-effect library) or 'clueso' (curated sfx) • image_type — images + provider='pixabay': 'photo' | 'illustration' | 'vector' • video_type — videos + provider='pixabay': 'film' | 'animation' • id — exact id; returns just that one record (any type) • source — media only: 'org' | 'stock' | 'all' (default = org + stock). Under 'all', stock is appended only when a query is given. sfx is stock only. • folder_id — projects + saved media (images/videos/music): restrict to a folder • engine / language — voices only • creator_id / mine_only — clueprints only • orientation — stock images/videos: 'landscape' | 'portrait' | 'square' • color — stock images: a color name/hex, e.g. 'blue' • size — stock videos: 'large' | 'medium' | 'small' • min_duration / max_duration — stock videos + freesound audio: length bounds in seconds • page / limit — paging for large sets (projects, components, clueprints — a clueprint list is sliced to the limit with no marker when more exist, so page through rather than assuming the first page is everything); stock media ignores these (fixed shortlist) Returns { type, count, items: [{ id, name, type, ... }] }. Feed the returned id straight into the consuming tool (set_voice, update_clips background, generate_media style_id, add_audio src, use_clueprint, etc.). Any `duration` on a returned item is in SECONDS — pass it straight to add_audio's source_duration.

Remove an element, a group or an audio track from a project. - target="element": removes an element from a clip (requires clip_index + element_id) - target="group": removes a GROUP node (requires clip_index + group_id). By default its children survive — they rise to the removed group's own parent, which for a top-level group is the clip root. Pass keep_children=false to delete the whole subtree instead, every nested group and every element inside it. - target="audio": removes a music/SFX track (requires music_id — returned by add_audio) Concurrency: target='element' is element-scoped (conflict domain: the individual element) — parallel-safe with other element edits on different elements, same as remove_elements. target='audio' is a whole-project mutation — serialize it against any other mutation on the same project_id. (Mutations to different projects run in parallel freely.)

Generate an AI image or canvas-code-based animation directly into a clip. - kind="image": text-to-image. Pass `prompt`. Optional: `animation_setting` (entry/exit — set it HERE, see below), `style_id` (from find type='image_gen_style_packs'), `reference_image_url` or `mcp_upload_id` for image-to-image grounding. - kind="animation": canvas-code animation rendered from a prompt. Pass `prompt`. Optional: `voiceover_text` (drives timing), `base_component_id` (reuse a saved animation as the starting point), `reference_image_url` or `mcp_upload_id` for visual grounding. Generation is asynchronous: the element is created immediately with a stable `element_id` and rendered in the background. Poll `get_clip(select:['busy'])` — an EMPTY `busy` means the render has landed. (This previously said to watch the `phantom` flag; `phantom` has never been a key get_clip returns, so there was nothing to poll.) Set presentation up front. `animation_setting` is applied to the element as it is created, so the image enters correctly the first time it renders. Doing it afterwards with `update_elements` means writing to the element that is still generating, which is the write most likely to be refused while the generation holds it. `group` is NOT accepted here, unlike `add_elements`: a generated element is built in the background, and the grouping would be overwritten when the render lands. Add it ungrouped, then call `update_elements` with `group` once it appears. Tip: use this tool whenever the user asks for a "generated", "AI", or "create me a" visual. For uploaded photos / logos / icons / GIFs, use `add_elements` with `element_type='image'` and a `src` or `mcp_upload_id` instead.

Transform, keyframe, rename, nest or lock GROUPS. A group moves, scales, rotates and fades as ONE unit, without touching any child's stored position — which is how you get a camera push instead of building the target layout twice and crossfading it. Address by `group_id`, never by name. Ids come from get_clip, or from the `group_id` that add_elements/update_elements hands back when you pass `group`. An unknown id is an error, never a silent create. There is no add_groups: a group is CREATED by putting elements in one — `add_elements(group: "hero-card")` — which also returns its id. That is not a shortcut; a group's pivot comes from its children's bounding box, so an empty group has no geometry to transform. The shortest path, in one script: const clip = get_clip({ project_id, clip_index: 2, select: ["groups"] }); const byName = new Map(clip.groups.map(g => [g.name, g.id])); update_groups({ project_id, items: [{ clip_index: 2, group_id: byName.get("hero-card"), anchor_px: [960, 540], keyframes: [ { timestamp: 0, scaleX: 1, scaleY: 1 }, { timestamp: 1.2, scaleX: 2.4, scaleY: 2.4, offsetX: -400, offsetY: -120, interpolation: "easeOut" }, ], }]}); Units are canvas pixels, as everywhere else: `offset_px`, `blur_px` and `anchor_px` are pixels; `scale` is a multiplier; `rotation`/`rotate` are degrees; `opacity` is 0-100. `anchor_px` is the pivot everything rotates and scales about. Give it as a point on the canvas — the server converts it to a fraction of the group's child bounding box and reports back both that fraction and the box, so you can check it. Default is the box's centre. Set it when you want a push that favours one corner. Only fields you pass are changed. `keyframes: []` CLEARS every track; omitting `keyframes` leaves them alone; a non-empty list replaces only the properties it names. `parent_group_id: ""` moves a group to the clip root; omitting it leaves the parent alone. Keyframable: offsetX, offsetY, scaleX, scaleY, rotation, rotateX, rotateY, opacity, blur. Position, size and colour belong to elements, not groups. Concurrency: clip-scoped (a group edit can add or re-parent nodes in its clip). Serialize against other mutations on the SAME clip; different clips and different projects run in parallel freely.

Create a screen-recording clip in a project. Creates blank placeholder clips, registers job entities, and sends the job to AVS. The blank clips this tool creates are placeholders; they become video clips when processing completes, so removing one loses that scene. Article placeholders are also inserted automatically into plainDoc. Requires the Auto-Recording add-on and per-workspace sign-in credentials for the product being recorded. Workspaces without it get back the manual path instead (upload_file, then add_clips(kind='video')) rather than a failure.

Reorder the elements inside a clip. List position IS the z-index — later entries in `ordered_element_ids` render on top of earlier ones. You MUST pass the full set of element IDs currently in the clip. The handler rejects partial lists so a reorder can never silently drop an element. Get the current list via get_clip — the order it returns is exactly the order this takes. Works on grouped clips. Elements keep their group; moving a group's members together moves the whole group relative to everything else. The one thing a group cannot do is render in two pieces, so an order that puts a non-member BETWEEN two members of the same group is rejected and names the group — put the group's elements next to each other, or take the element out of the group first with update_elements(group).

Split a video clip into two clips at the given timestamp. The left clip keeps the original clip ID; the right clip gets a new ID and is inserted immediately after. Elements spanning the split point are duplicated into both clips with adjusted timing. Voiceover transcript and transcribed words are split proportionally. Concurrency: whole-project mutation (conflict domain: the entire project) — serialize. Do not run it in parallel with ANY other mutation on the same project_id, including element/voiceover edits; run them one at a time. (Mutations to different projects run in parallel freely.)

Switch the active workspace. The choice sticks for this client until you change it again — including across reconnects — so every later call operates in the new workspace. Use find(type='workspaces') to see available options.

Insert one sync marker on a clip's transcript. Use this when: - The user is explicit about WHERE the camera should pause / cut (e.g. "sync the word 'submit' to 4.2s of the demo"). - `auto_sync` ran but missed a step you care about. How matching works: - `word`: case-insensitive, punctuation-stripped. The first match in the transcript is used unless `occurrence > 1`. - `occurrence`: 1-indexed — pass 2 to target the SECOND time that word appears, 3 for the third, etc. Required when the word repeats. - `timestamp_seconds`: clip-relative seconds. When the clip has run TTS already (`generated_timestamps` present), the server inverse-maps this to original-recording seconds automatically. Constraints: the clip MUST be a video clip with a source recording (otherwise the frame thumbnail can't be extracted). The transcript must already contain the word — if not, you'll get `word_not_found` with a 200-char excerpt of the transcript to help you retry.

Upload one or more files to Clueso. Three modes — pick by client + where the file lives: 1. **file_name** — HOSTED upload, the default for any non-UI / programmatic upload (Claude Code, Cursor, Claude Desktop, scripts). Returns an upload URL on Clueso's OWN base domain + a ready-to-run curl that streams a single local file to it; Clueso relays the bytes to storage server-side. The PUT targets the base domain — NOT cloud storage directly — so it works on desktop/agent clients that can't reach or are blocked from S3. Requirement: the client must be able to PUT bytes to the Clueso base domain (run the returned curl, or any HTTP PUT). The agent (or the user at a shell prompt) runs the curl. Prefer this whenever there's no human at a browser. 2. **file_url**: Pass a public https URL. Server fetches and stages the file. Returns mcp_upload_id immediately. Use when the file is already on the open web — no user interaction needed. 3. **request_hosted_upload** (UI mode — use ONLY when a human should pick files in a browser: many files at once, or a host with no shell / no PUT capability): Returns a single upload_token + upload_page URL. Share the link with the user; they open it in a new browser tab, drop their files, click Done. Then call check_uploads(upload_token) to retrieve all mcp_upload_ids. Call once for all files. Hosted uploads cover any number of files per call: one call issues one upload_token, and that token covers every file the user drops on the page. Repeat calls issue additional tokens, each tracking only its own files. The returned mcp_upload_id (prefixed `mup_`) can be passed to: - add_elements / update_elements (image or video → an element ON a clip: pass it as `type_data.mcp_upload_id`, on either tool — this is how a local image becomes on-canvas content, and how an existing element's source is swapped). To fill an animation's image slot, pass it inside `type_data.parameter_values` on update_elements only — parameter_values is an update-path field and is stripped on add. - add_audio (audio → project music track that plays under all clips) - add_clips(kind='video') (video or audio → sequential clip with auto-transcription) - add_clips(kind='pptx') (.ppt/.pptx → slide clips) - add_article_media (image/GIF → article asset) - analyze_audio (audio → transcript / silences / beats / features)

Fetch all files the user uploaded via the hosted upload page. Returns one mcp_upload_id per file — pass each to the appropriate tool (add_clips, add_article_media, etc.). Call this once after the user confirms they are done uploading; the token covers all files they dropped on the page.

Set the voiceover voice for every clip in a project. Use this to apply a clueprint's voice (read `voiceover.voice.name` and `voiceover.voice.engine` from the clueprint source data), or to switch all clips to a specific voice in one call. The voice is looked up by name + engine; lookup is case-insensitive on the name. Common engines: 'eleven' (ElevenLabs), 'cartesia', 'google'.

Set voiceover text and/or trigger speech generation for any number of clips in a single call. Each entry chooses its own action: - "set_text" — set transcript for a clip (clip_index + text) - "generate_speech" — async TTS for a clip (returns immediately) - "set_and_generate" — set text and kick off TTS in one entry (+ text) Address a clip the same way as everywhere else: `clip_index`. Pass `clip_id` instead if you already have it — it survives clips being inserted or reordered mid-build — but you never need both; whichever you omit is looked up once for the whole call. Entries within one call are applied in order. Returns one result object per input entry. All text-set actions land in ONE save; the TTS for generate/set_and_generate runs async per clip after. IMPORTANT — generating speech RESCALES the whole clip, it does not clamp it: when audio is generated (generate_speech / set_and_generate), the clip's duration is reset to the spoken audio length, and then EVERY element on that clip is retimed proportionally by (new duration ÷ old duration). start_time, end_time and every keyframe timestamp are multiplied by that factor. Nothing is merely truncated — on a 6s clip that becomes 1.02s, an animation you placed at [0, 1.6] ends up at [0, 0.27]. Zoom elements whose window falls under the minimum after scaling are DROPPED entirely. Generation is async, so this lands AFTER this call has already returned success. So: generate speech BEFORE placing time-sensitive elements, or size them against estimate_duration first — then re-read the clip and check what your elements actually became, not just the clip duration. Concurrency: parallel-safe (conflict domain: a clip's voiceover). The server merges each clip's voiceover under a per-guide lock and preserves that clip's elements, so you can fan voiceover work out across subagents by clip — and it's safe to run alongside element edits. Two concurrent edits to the SAME clip's voiceover do not last-write-win — both claim that clip's voiceover path, so the later one is REJECTED and nothing is written; re-read and re-apply. Do NOT run concurrently with whole-clip/whole-project mutations on the same guide (update_clips on that clip, structural clip ops, add_audio, update_project).