MCP tools reference

Voice synthesis and transcription run on the user's own provider key, and a brand kit is established once so every later video inherits the same colours, fonts and logo. These tools are called before the first video rather than per video, and they are the only ones that touch account-level state instead of a job.

set_voice_key

Store the user's own ElevenLabs or Cartesia API key, used for both narration and transcription. Required before approving a narrated or recorded video on a non-allowlisted account; stored encrypted and never returned.

get_voice_status

Report whether a voice key is connected, which provider it belongs to, and which narration engine that provider will use. Check it before approve so a missing key is caught early.

clear_voice_key

Delete the saved voice provider key. After this, a non-allowlisted account must call set_voice_key again before a narrated video can be produced.

set_brand_kit

Establish the user's brand from their website URL — colours, fonts, logo, name and tone are read deterministically from the page, with no model involved. Any field can be overridden explicitly.

get_brand_kit

Read the stored brand kit that every new video inherits. Returns null when none is set, which is the cue to offer set_brand_kit with the user's website.

clear_brand_kit

Remove the brand kit and its stored logo. Videos already made keep the look they were rendered with; new ones fall back to the unbranded looks.

get_niche_profile

Return the user's saved content profile — niche, sub-topics, audience, goals, content pillars and topics to avoid — so a vague brief can be turned into a concrete topic that matches what they publish.

A video is created as one of three kinds: an explainer with a narrated script, a silent motion graphic driven by on-screen text, or a recording the user uploads and the assistant cuts. Aspect ratio and duration are independent axes, so any length works at either 9:16 or 16:9.

get_creation_options

Open a picker where the user chooses the look and the narrator voice themselves, with playable preview clips and voice samples, and creates the job from that panel. Clients that cannot render panels receive the same options as text.

create_job

Start a brand-new video job with a kind (explainer, motion_graphic, recording), a topic, an aspect ratio, a duration and an optional genre style. Returns a playbook plus the first brief for the chosen kind.

Nothing is produced until the user approves a plan, and the plan's shape follows the video's kind: a script for an explainer, a shot list for a motion graphic, or a keep-and-drop cut plan for an uploaded recording. The same three tools serve all three, so a new kind of video adds no new tools.

get_plan

Fetch the plan to write before production — the script brief, the shot-list brief, or an uploaded recording's transcript split into keep/remove segments. After approval it returns the plan the video is currently running, with real per-shot durations.

submit_plan

Submit the script, shot list or cut plan for the user's approval, replacing any pending plan. It also retimes a finished motion graphic: resubmit the same shots with different durations to make a scene or the whole video shorter.

approve

The user's go button. Starts production for an explainer or motion graphic, or confirms the cut plan for a recording. Only call it after the user has explicitly confirmed, in chat or in the review panel.

This is where the assistant does the creative engineering. It writes a direction for each scene, then a self-contained Remotion component for each scene, and renders stills to check the result with its own eyes. Submitted code is statically validated and executed in a sandbox, so a crash comes back as a tool error to fix rather than a broken video.

get_direction_brief

Fetch the creative contract for directing scenes, or one scene's brief including the previous scene's context so consecutive scenes are made to differ. Direct scenes one at a time, in order.

submit_directions

Submit one scene's direction — its mood, layout, beats and optional stage framing. Submissions accumulate; when every scene is covered the batch finalizes and the job advances to code generation.

get_code_brief

Fetch the full Remotion engineering contract — available globals, the resolved theme, the layout rules and the animation kit — plus one scene's spec. No access to this codebase is needed to write a scene.

submit_scene_code

Submit one scene's Remotion component. It is statically validated, smoke-rendered in a sandbox and audited for layout, and accepted scenes are assembled into the finished composition when the last one lands.

list_scenes

List every scene with its id, name, type, narration text, start and end times, duration and whether it already has accepted code. This is how a user's description of a scene is matched to a scene id.

get_scene_code

Read the Remotion component currently accepted for one scene, so a requested change edits the real code instead of rewriting the scene from scratch. Also lists that scene's hand edits made on the platform.

edit_scene_code

Apply exact-substring replacements to a scene's current code — the token-cheap way to make a small change. Each edit runs the same validation and sandbox render as a full submission and recomposes the video in place.

render_scene_preview

Render one scene to still images on the server and return them, so the assistant can see exactly what it built without asking the user to screenshot anything. Captions, hand edits and the speaker's footage are composited in.

A video can draw on the user's own files, on a hosted library of CC0 music, and on real brand logos and stock photography resolved through the same providers the platform's own director uses. Uploads go straight to object storage with a presigned link rather than through the API.

attach_media

Attach the user's own media to a job in one of three roles: a reference image or clip to embed in a scene, the primary recording for an upload-and-animate job, or a background music track from the platform's CC0 library.

list_media

List every asset attached to a job with its id, kind, dimensions, caption and the exact URL to embed in scene code. Images come back as inline image content so the assistant can tell them apart.

resolve_brand_asset

Resolve a real company or product logo to a stable URL that a scene can embed. Returns null rather than a guess when no logo is found, which is the signal to draw the mark as inline SVG instead.

resolve_scene_image

Resolve a stock photograph for a concrete real-world subject and return a stable URL plus its attribution. At most one image per scene, and never for abstract concepts, which are better served by motion graphics.

These three tools change a finished video without touching any scene code, which makes them cheap, reversible and safe to run on a video the user has already exported. Each is deliberately structured rather than conversational, because the assistant cannot hear or see the result and has to confirm from what comes back.

adjust_audio

Mute or rebalance everything except the narration — the generated sound effects and the user's background music — scoped to one cue, one scene or the whole video. Calling it with only a job id reads back the full cue list.

adjust_captions

Turn burned-in captions on or off, switch preset, font, size, colours or placement, or correct a mis-heard word by its transcript index without re-recording anything. Narrated videos only; a silent motion graphic has none.

adjust_scene_framing

Re-frame the speaker in one scene of a recorded video: which part of the footage shows, where it sits on the canvas, and when that layout changes partway through the scene. Framing changes do not invalidate the scene's code.

Narrated and recorded videos carry a publishing kit — titles, descriptions, hashtags, tags and chapter markers — written by the assistant rather than by a server-side model. Submitting it is what unblocks the platform's export action for those kinds; silent motion graphics skip the step entirely.

get_social_brief

Fetch the strategist contract for this video's social posts: the per-platform format rules, the topic, the full approved script and, for widescreen videos, the exact chapter timestamps.

submit_social_metadata

Persist the YouTube and Instagram copy written from the brief. Once accepted it appears in the user's Share panel and the platform's export action is unblocked. Resubmitting overwrites what is stored.

Nothing reports progress on the assistant's behalf, so a long-running stage is polled. Recovering a job id later in a conversation matters more than it sounds: creating a second job to fix an existing video is the single most common way an agent produces a duplicate instead of an edit.

get_job_status

Return the current pipeline stage, progress, any error, guidance naming the next tool to call, and — once a render finishes — the download URL. Poll it roughly every fifteen seconds while production runs.

list_jobs

List the user's most recent jobs newest first, each with its id, topic, format, stage and web URL. This is how a job id is recovered when the user asks to change a video made earlier in the conversation.