analyze_audio
ChatGPTMeasure one audio file or a video’s audio track. TAKES IN: an asset UUID, an optional window in milliseconds (at most 300 seconds analyzed) and optional report sections. PRODUCES: measured loudness, dynamics, clipping, silence, spectral, rhythm and section reports plus credits consumed. This is paid signal analysis, not speech transcription or a subjective quality score. Source media is unchanged. Failed analysis is refunded.
analyze_video
ChatGPTInspect one stored video for what is on screen: text, objects, framing, visible defects, adherence to a brief. TAKES IN: a video asset UUID (at most 120 seconds long) and a specific question. PRODUCES: a text interpretation and credits consumed, no new asset. A paid model analysis, not a guarantee; source unchanged; a failed or empty analysis is refunded. It samples frames, so it misses motion under a second (cut or crossfade, a slow zoom, exact cut times) yet may still answer confidently.
audio_compose
ChatGPTJoin audio clips end to end into one audio file, such as narration or music segments, with silence in any gaps. TAKES IN: one or more audio asset UUIDs placed on a timeline. PRODUCES: one NEW MP3; sources are unchanged. Clips must not overlap. Gaps become silence. This assembles narration or music segments; video_compose adds audio to a video. Returns generation_ids and spends credits.
audio_compress
ChatGPTEven out the loud and quiet parts of one audio file with a compressor, as a new audio file. TAKES IN: one audio asset UUID, threshold_db and ratio, and optional attack_ms, release_ms and makeup_db. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_deess
ChatGPTSoften the harsh "s" sounds of speech in one audio file, as a new audio file. TAKES IN: one audio asset UUID and intensity. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_denoise
ChatGPTReduce steady background noise in one audio file, as a new audio file. TAKES IN: one audio asset UUID and denoise_strength. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_equalize
ChatGPTBoost, cut or filter one frequency band of one audio file, as a new audio file. TAKES IN: one audio asset UUID, eq_kind, frequency_hz, and gain_db for the kinds that boost or cut. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_fade
ChatGPTFade one audio file in from silence, out to silence, or both, as a new audio file. TAKES IN: one audio asset UUID and fade_in_ms, fade_out_ms or both. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_gain
ChatGPTMake one audio file louder or quieter by a fixed gain, as a new audio file. TAKES IN: one audio asset UUID and gain_db. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_gate
ChatGPTQuieten everything below a level in one audio file with a noise gate, as a new audio file. TAKES IN: one audio asset UUID and threshold_db, and optional ratio, attack_ms and release_ms. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_limit
ChatGPTHold the peaks of one audio file under a ceiling with a limiter, as a new audio file. TAKES IN: one audio asset UUID and ceiling_dbtp, and optional attack_ms and release_ms. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_reverb
ChatGPTPlace one audio file in a space by adding reverb, as a new audio file. TAKES IN: one audio asset UUID, reverb_preset and wet, and an optional pre_delay_ms. PRODUCES: exactly ONE new audio file; the source is unchanged. Returns generation_ids and spends credits per call.
audio_trim
ChatGPTCut one audio file to a time range, as a new audio file. To join clips, use audio_compose; to cut a video, sound and picture together, use video_compose. TAKES IN: one audio asset UUID, start_ms and end_ms. PRODUCES: one NEW audio file; the source is unchanged. Returns generation_ids and spends credits per call.
browse_library
ChatGPTThe library view's own read of the organization's library: one page under the view's sort and filters, each row with the caller's favourite, the tags the caller can see, its folder and whether the caller made it. The library view makes this call; list_assets is the listing to use for anything else. TAKES IN: list_assets' filters and cursor, plus sort (newest, oldest or recently_modified), favorites_only, created_by_me, source (generated or uploaded), folder_id beside its project_id, and up to 10 tag_ids an asset must all carry, or untagged. PRODUCES: one page of rows and next_cursor for the same sort and filters, null on the last; changes nothing.
burn_subtitles
ChatGPTBurn subtitles or captions permanently into a video (hardcoded) from an SRT, WebVTT or ASS/SSA subtitle asset, with styling. TAKES IN: one video asset UUID and one subtitle document asset UUID, plus caption styling. PRODUCES: one NEW captioned video; sources are unchanged. Use transcribe_media first when captions do not exist, or import_media_from_url for an .srt the user has. Its cues start with a speaker label ("speaker_0: "), burned as written; tell the user first. Returns generation_ids. Holds the FilmLab render minimum, then settles credits to actual render compute.
change_voice
ChatGPTRe-speak an existing recording in another voice, keeping its words, timing and delivery. TAKES IN: exactly one audio_asset_id (the recording) plus one voice_id from list_voices; optionally a model_name from list_models under STS. PRODUCES: exactly ONE new audio asset per call, the same words in the new voice; the recording is unchanged; the result is a generation_id for optional status checks with get_generation_status. Use this to change WHO is speaking. To read new words aloud instead, use generate_speech. The charge is worked out from the recording's length at the start, so a recording longer than the model takes is refused before anything is spent, naming the ceiling.
delete_assets
ChatGPTDelete library assets as Studio's delete does: they leave the library, their shares and timeline clips go for good, and restore_assets brings the assets back for 90 days. The library view makes this call when the user deletes. TAKES IN: up to 10 asset_ids. PRODUCES: the deleted ids and restorable_until; a call naming any id the library does not show deletes nothing.
edit_image
ChatGPTEdit, restyle or combine one or more of the organization's images under a text instruction with an image-to-image model, and place the result in a Morphic project. TAKES IN: a prompt describing the change, image_asset_ids (1 or more image asset ids from list_assets or a finished generation), character_id (from list_characters; counts as one reference and may stand alone), plus optional model_name (from list_models under the ITI feature), aspect_ratio, resolution, variations (1 to 4, default 1) and project_id. PRODUCES: one NEW image asset per variation, each its own generation and charge, started asynchronously; the source images are unchanged. The result carries every generation_id for optional status checks with wait_for_generations and the credits the starts will charge. Duplicate image ids count once. Each model takes a maximum number of distinct reference images; when the chosen one takes fewer than given, the refusal names the models that take that many. Omit aspect_ratio to keep the source's framing. Do not call edit_image again for the same request. To enlarge an image without changing it, use upscale_image. To erase or replace one region and keep the rest of the picture, use inpaint_image.
edit_video
ChatGPTChange what is in a video clip (swap, add or remove something and keep the rest, or restyle it), or continue it past its end, with one of the organization's video-to-video models, and place the result in a Morphic project. TAKES IN: mode (edit or extend), video_asset_id, and usually a prompt; optional model_name (from list_models under VTV), reference images, videos and audio, duration_seconds (required to extend), resolution, aspect_ratio, with_audio, bitrate_mode, character_orientation, subtitles and project_id. PRODUCES: exactly ONE new video asset per call, started asynchronously; the source is unchanged. The result carries the generation_id and the credits the start charges. Pick the model by the job: reframe to another shape (vertical, square, widescreen), Luma Ray 3.2 Reframe with aspect_ratio; move a subject image the way the video moves, Kling V3 Motion Control; captions, Veed Subtitles; sound effects, Sonilo SFX or Mirelo SFX 2.0; extend, Seedance or Morphic Extend; restyle or replace something in the scene, Kling O3, Gemini Omni Flash 1.1, Grok Imagine, Morphic Edit or Seedance. An edit keeps the source's length; for a new video that only draws on this one, use generate_video. Video takes one to several minutes; do not call edit_video again for the same request. State the estimated_credits to the user.
enhance_image
ChatGPTRepair an image without changing its size: remove noise, fix blur, restore an old photo, correct its lighting. TAKES IN: exactly one image_asset_id from the library; optionally a model_name from list_models under IE. PRODUCES: exactly ONE new image asset per call, the same size as the source; the source is unchanged; the result is a generation_id for optional status checks with get_generation_status. Each model fixes a different fault and list_models says which, so name the model that matches the problem. To make an image LARGER, use upscale_image; to change what is in the picture, use edit_image.
enhance_video
ChatGPTRepair a video without changing its size: remove noise, fix blur, convert it to HDR. TAKES IN: exactly one video_asset_id from the library; optionally a model_name from list_models under VE. PRODUCES: exactly ONE new video asset per call, the same size as the source; the source is unchanged; the result is a generation_id for optional status checks with get_generation_status. Each model fixes a different fault and list_models says which, so name the model that matches the problem. To make a video LARGER, use upscale_video. The charge is worked out from the video's length and its size, and a video longer than the model takes is refused before anything is spent.
favorite_assets
ChatGPTStar or unstar library assets for the caller alone, as Studio's favourites do. The library view makes this call. TAKES IN: up to 10 asset_ids and favorite, true or false. PRODUCES: the ids and their state; a call naming any id the library does not show changes nothing.
finalize_video_draft
ChatGPTRender a Seedance 2.5 Draft at 1080p: the final of a draft the organization made in the last 7 days, placed in a Morphic project. TAKES IN: the draft's video_asset_id and an optional project_id. PRODUCES: exactly ONE new 1080p video asset, started asynchronously; the draft is unchanged. The result carries the generation_id and the credits the start charges, per second of the draft. A draft's asset carries final_generation_expires_at; past it, only a new draft can be finalized. Only Seedance 2.5 Draft outputs qualify; to enlarge any other video, use upscale_video. Do not call it again for the same draft unless the user wants a second final.
generate_3d_model
ChatGPTGenerate one textured 3D model (a GLB file) from images of an object, and place it in a Morphic project. A description alone is not enough: to model an object you only have in words, generate an image of it first. TAKES IN: reference_image_ids (1 to 5 image assets showing the object, required), an optional prompt describing it, plus optional model_name (from list_models under the TI3D feature) and project_id (from list_projects). PRODUCES: exactly ONE new 3D model asset per call, a GLB, started asynchronously; the result carries the generation_id for optional status checks with get_generation_status, whose finished asset url downloads the GLB, and the credits the start will charge. A 3D model takes a few minutes; do not call generate_3d_model again for the same request. Images of one object from several sides give the most faithful shape.
generate_dialogue
ChatGPTTurn a script of several speakers into one audio file (text to dialogue). TAKES IN: 2 to 10 lines in order, each with its text and either a voice_id from list_voices (the catalog models) or a short speaker description (a model that reads the cast from the text, like "Seed Audio 1.0"), plus optionally a model_name from list_models under TTD. PRODUCES: exactly ONE new audio asset per call with every line spoken in order; the result is a generation_id for optional status checks with get_generation_status. Every voice_id in one dialogue must belong to the chosen model's catalog; each model's own script length applies, and some models cap the number of distinct voices, both named in the refusal. For a single speaker use generate_speech.
generate_image
ChatGPTGenerate one or more images from a text prompt with one of the organization's models and place them in a Morphic project. TAKES IN: a prompt, plus optional model_name (from list_models under the TTI feature), aspect_ratio, resolution, variations (1 to 4, default 1) and project_id (from list_projects). PRODUCES: one NEW image asset per variation, each its own generation and its own charge, started asynchronously; the result carries every generation_id for optional status checks with wait_for_generations and the credits the starts will charge. Generation takes seconds to a minute; do not call generate_image again for the same request. To change an existing image, use edit_image; to enlarge one, use upscale_image.
generate_music
ChatGPTCompose a piece of music from a description (text to music). TAKES IN: a prompt describing the music, and optionally lyrics, instrumental (no vocals), a length in seconds and a model_name from list_models under TTM. PRODUCES: exactly ONE new audio asset per call; the result is a generation_id for optional status checks with get_generation_status. Each model has its own length range or fixed set, named in the refusal when the requested length is unsupported; omit duration_seconds for the model's default. credits_settle_on_output_duration on the result says whether the charge follows the produced length or the requested one.
generate_sound_effect
ChatGPTMake a sound effect from a description (text to sound effects): ambience, foley, impacts, transitions. TAKES IN: a prompt describing the sound, and optionally a length in seconds and a model_name from list_models under TTSE. PRODUCES: exactly ONE new audio asset per call; the result is a generation_id for optional status checks with get_generation_status. Each model has its own length range, named in the refusal when the requested length is unsupported; omit duration_seconds for the model's default. Use generate_music for music and generate_speech for spoken words.
generate_speech
ChatGPTTurn a text into spoken audio in a chosen voice (text to speech). TAKES IN: the text to speak, a voice_id from list_voices for the same model, and optionally a speech model_name (from list_models under TTS) and a language. PRODUCES: exactly ONE new audio asset per call, spoken as written; the result is a generation_id for optional status checks with get_generation_status. The catalog models need a voice_id; "Seed Audio 1.0" takes none and reads the voice from the text itself (age, tone, accent, pace). Each model has its own text length limit and the refusal names it. estimated_credits is the whole charge unless credits_settle_on_output_duration is true; then it is the floor, settled when the audio is finished. Do not call again for the same text while a generation is running.
generate_video
ChatGPTGenerate a video with one of the organization's video models, and place it in a Morphic project: from a prompt alone, from one image animated, from references the model composes a scene out of, between a first and a last frame, or through keyframes at set times. TAKES IN: a prompt, plus AT MOST ONE of image_asset_id (one image to animate, ITV), the reference_* fields (image, video and audio references and one document, RTV), frame_image_asset_ids (exactly two images, the first frame and the last frame in that order, FIV) or keyframes (images with a time each, FIV); a prompt with none of them is text to video (TTV). Also optional: model_name (from list_models under whichever of those features the sources pick), duration_seconds, resolution, aspect_ratio, with_audio, bitrate_mode, prompt_expansion_mode and project_id. PRODUCES: exactly ONE new video asset per call, started asynchronously; the result carries the generation_id for optional status checks with get_generation_status and the credits the start will charge. References are numbered per kind in the order given: Seedance models read @Image1, @Video1, @Audio1 in the prompt; MiniMax models read "Image 1", "Video 1". A reference video is material for a new video; to change a video itself (restyle, reframe, caption, extend), use edit_video. Video takes one to several minutes; do not call generate_video again for the same request. Video costs many times more credits than an image; state the estimated_credits to the user. To enlarge an existing video, use upscale_video.
get_credit_balance
ChatGPTReport the connected organization's plan and the credits it can still spend. TAKES IN: nothing. PRODUCES: the plan name, whether credits are unlimited, remaining_credits (the subscription pool plus the purchased pool, capped by the caller's personal allocation when the organization set one) and each pool's used and total, read-only. Budget expensive work such as video from list_models, which carries each model's credit price per resolution: a generation tool's own estimated_credits arrives with the start it reports, so it cannot gate that start.
get_generation_status
ChatGPTOnly for a generation a start returned, by its generation_id: its status and finished assets; it displays nothing. A library asset's details come from list_assets. TAKES IN: generation_id and optional wait_seconds (0 to 45) to wait until it finishes or time runs out. PRODUCES: status, error_message, credits_charged and assets with public URLs. Read-only. Read status only when the server instructions say to follow a start, the user asks for status or a link, or a requested next step needs the output; prefer wait_for_generations for those checks. Use show_generation only to display an earlier result on request.
get_workflow
ChatGPTRead one saved workflow in full: the steps it follows and the inputs it expects. TAKES IN: exactly one workflow_id from list_workflows. PRODUCES: the workflow's active version — its steps as its author wrote them, its stages, and the inputs a run collects first. A workflow another organization owns is listed but its steps are withheld: steps and stages come back empty and shared_from_another_organization is true, while the inputs stay complete. Carry out the steps with the generation tools; this server does not execute a workflow for you. Ask the user for every required input before you start.
image_compose
ChatGPTCombine several images into one image: side by side in a row, stacked in a column, or in a grid (a collage). TAKES IN: ordered image asset UUIDs and layout settings. PRODUCES: one NEW image; sources are unchanged. Choose horizontal, vertical or grid layout. Grid requires columns. Fit scale resizes images; pad retains their size within slots. Returns generation_ids and spends credits.
image_crop
ChatGPTCrop a region out of an image, or cut it into a grid of equal tiles, each a new image. TAKES IN: one image asset UUID and normalized crop bounds, optionally a grid. PRODUCES: one NEW image per cell, or one crop without a grid; the source is unchanged. Bounds run from 0 to 1. Each output has its own generation_id; the call costs 1 credit per 2 cells, rounded up.
image_extend_to_aspect_ratio
ChatGPTPad one image to a new shape with white, without an AI model; nothing is drawn in the new area. To fill it with generated scenery, use outpaint_image; to cut a region, image_crop. TAKES IN: one image asset UUID, aspect_ratio and optional anchor. PRODUCES: one NEW image; the source is unchanged. Returns generation_ids and spends credits per call.
image_flip
ChatGPTMirror one image left to right or top to bottom, without an AI model. TAKES IN: one image asset UUID and axis. PRODUCES: one NEW image; the source is unchanged. Returns generation_ids and spends credits per call.
image_make_color_transparent
ChatGPTMake one flat background colour of an image transparent, without an AI model, so the image can sit over others (image_compose). TAKES IN: one image asset UUID, color (or "auto") and optional tolerance. PRODUCES: one NEW image with transparency; the source is unchanged. Returns generation_ids and spends credits per call.
image_resize
ChatGPTResize one image without an AI model: it fits inside the size you give, or is stretched to it when maintain_aspect is false; what it shows is unchanged. To change the content, use edit_image. TAKES IN: one image asset UUID and a width, a height or both. PRODUCES: one NEW image; the source is unchanged. Returns generation_ids and spends credits per call.
image_rotate
ChatGPTRotate one image clockwise by a quarter, half or three-quarter turn, without an AI model. TAKES IN: one image asset UUID and degrees (90, 180 or 270). PRODUCES: one NEW image; the source is unchanged. Returns generation_ids and spends credits per call.
import_media_from_url
ChatGPTBring an image, video or audio file, or a document (.pdf, .txt, .md, .srt), from a public HTTPS URL or a chat attachment into the organization's assets library. TAKES IN: either url, an HTTPS URL that serves the file itself (not a web page), or file, a file the user attached that the host hands over as a file object (download_url, file_id); an optional name and an optional project_id from list_projects. PRODUCES: exactly ONE new library asset per call, uploaded as if from Morphic Studio, with the id the generation tools take (image_asset_id, video_asset_id, a document id) and its URL; the source is unchanged. A document's contents are never returned. A signed import_url from prepare_media_upload is supported after the client completes its PUT upload; the object need not have public read access. The host must be a public domain name (IP addresses and local hosts are refused), the response must declare an image, video or audio content type or the URL must name a document, and the file must fit the size cap named in the refusal. A PDF must open without a password; a text document must be UTF-8, and an .srt must hold SubRip cues. Redirects are followed a few hops, each checked the same way. Calling twice with the same URL adds two assets. Use list_assets to find a file already in the library instead of importing it again.
inpaint_image
ChatGPTReplace one region of an image with something else, keeping the rest untouched. TAKES IN: exactly one image_asset_id, a prompt for what belongs in the region, and that region as a box in the pixels of the source itself (x, y, width, height from its top left corner); optionally a model_name from list_models under IPG. PRODUCES: exactly ONE new image asset per call, the source with that box repainted; the source is unchanged; the result is a generation_id for optional status checks with get_generation_status. Read the source's width and height from list_assets or from the generation that made it, and give a box that lies inside them. Erasing something is this tool: name the box it sits in and say what belongs there instead. The box is filled from the prompt alone, so to place a SPECIFIC image into a spot, use edit_image. To change the whole picture instead, use edit_image; to add room around it, use outpaint_image.
lip_sync
ChatGPTMake a person in a video or a still image speak an audio track (lip sync). TAKES IN: exactly one of video_asset_id (a video whose speaker is re-voiced) or image_asset_id (a still that is animated), plus audio_asset_id (the speech to follow), all library ids; optionally a prompt, a model_name from list_models under LS and a resolution. PRODUCES: exactly ONE new video asset per call with the audio track; the sources are unchanged; the result is a generation_id for optional status checks with get_generation_status. The video-driven models keep the video's length; the image-driven ones run for the audio's length. Each model takes either a video or an image beside the audio, named in the refusal when the given kind does not fit; omit model_name to get the default model for the kind given. A model that sizes its output from the source takes no resolution.
list_asset_tags
ChatGPTThe organization's tags the caller can see, its public ones and the caller's own private ones, by name, for the library view's tag filter and tagging. The library view makes this call. TAKES IN: nothing. PRODUCES: each tag's id, name, color, visibility and whether the caller created it; changes nothing.
list_assets
ChatGPTList the connected organization's library: stored uploads and finished generations, newest first, with public URLs. TAKES IN: optional type (image, video, audio or document), project_id (from list_projects), a case-insensitive name_contains filter and a limit. PRODUCES: the matching assets with id, type, name, size, dimensions, duration or pages, URL, source, project and the generation that produced them, read-only; never a document's contents. Use a returned id wherever another tool asks for an asset: edit_image, upscale_image, upscale_video, or generate_video to animate an image.
list_characters
ChatGPTList the characters this organization has trained, so a later call can put one in a picture. TAKES IN: nothing, or a name filter and a page size. PRODUCES: the characters that finished training, newest first, each with the character_id edit_image takes. A character is a likeness the organization trained in Morphic Studio; this server cannot train one. Characters another organization shared, including ones shared with every organization, are listed too and marked as such.
list_elements
ChatGPTList the Elements this organization can use: the characters and styles it trained, so a later call can put them in an image or a video. TAKES IN: nothing, or a kind (character or style), a name filter, a page size and a cursor. PRODUCES: one page of the Elements that finished training and are not locked, newest first, each with its kind and the element_id a generation tool takes in element_ids, and a next_cursor when more remain. An Element is a likeness or a look the organization trained in Morphic Studio; this server cannot train one. Elements another organization shared, including ones shared with every organization, are listed too and marked as such. A character_id from list_characters is the same id.
list_generations
ChatGPTList the connected organization's recent generations, newest first, each with its status and finished assets. TAKES IN: an optional generation feature code to keep one kind, and a limit. PRODUCES: the generations with generation_id, feature, prompt, model_name, created_at, status, error_message and assets with public URLs, read-only. A long prompt is cut to its head, marked with its full length. Generations made in Morphic Studio and by the Morphic copilot appear too, not only the ones this client started. After a start whose response was lost, look here before retrying it: a repeated start is charged again.
list_models
ChatGPTList the connected organization's generation models and prices. TAKES IN: optional feature code and model_name. PRODUCES: models grouped by feature, with best_for, availability, pricing and the MCP inputs that can run each; read-only. An unavailable model cannot be used; unavailable_reason explains the restriction. pricing_type names the base rate. Add reference charges per the billing fields; null rates mean unsupported. Starts return estimated_credits; completed status returns credits_charged including settlement and refunds. mcp_bindings lists the tool/input combinations that can run each model, with the inputs that model takes; an empty list means the model exists in Morphic but is not exposed by MCP. A named model_name also returns its capabilities: duration choices/ranges, aspect ratios, reference counts and input-duration bounds, audio controls, upscale factors/FPS, org-allowed resolutions, and the model's own options with their allowed values. Model capabilities may be broader than each binding's input fields.
list_project_folders
ChatGPTOne project's folders, for the library view's folder filter and its move to a folder. The library view makes this call. TAKES IN: project_id and the previous page's next_offset. PRODUCES: up to 100 folders with id and name, and next_offset, null on the last page; changes nothing.
list_projects
ChatGPTList the connected organization's projects, newest updated first. TAKES IN: an optional limit and an optional case-insensitive name_contains filter. PRODUCES: the matching projects with id, name, description, status and timestamps, read-only. Use a returned id wherever another tool asks for a project.
list_voices
ChatGPTList the voices a speech model can speak with, from its provider's catalog. TAKES IN: an optional speech model_name (from list_models under TTS; the default speech model when omitted), a language, and optional name, gender and accent filters. PRODUCES: one page of voices, each with the voice_id that generate_speech takes for the same model. No credits are charged, but each call counts toward the connection's hourly limit on imports and voice listings, and the voices it lists are recorded as reusable ids. A voice belongs to one provider's catalog, so it works with the models of that provider only; list again with the other model_name to speak with another model. Gender and accent filters apply to the ElevenLabs catalog; the other catalogs are small and name their style or native language instead. Accent labels match exactly, ignoring case; pages can be short or empty after filtering. has_more says whether a next provider page exists.
list_workflows
ChatGPTList the saved workflows this organization can run: the recipes its people wrote for work they repeat. TAKES IN: nothing, or a name filter and a page size. PRODUCES: the workflows visible to the organization, most recently updated first, each with the workflow_id get_workflow takes. A workflow holds the steps, not a button: read one with get_workflow and carry out its steps with the generation tools. A row marked shared_from_another_organization can be run from its inputs, but get_workflow withholds its steps.
morphic_3d_create_scene
ChatGPTCreate a new 3D scene in a project, empty or from a whole scene, for example one rebuilt from a Blender file. TAKES IN: name, optional project_id, optional scene in the shape morphic_3d_read_scene returns. PRODUCES: exactly one new scene and its scene_id. Charges no credits. Publish the scene's models first (morphic_3d_prepare_model_upload, morphic_3d_publish_model) and put each model_ref in its object's assetId. A scene the engine refuses is not created, and the refusal names the field.
morphic_3d_get_model_download
ChatGPTGet a download link for one 3D model's file, for example to import it into Blender. TAKES IN: model_ref. PRODUCES: download_url for the main file (file_name says its type, .glb, .gltf or .obj), extra_files with a download_url for every file the main file reads (empty for a self-contained file), and preview_url; it changes nothing. Download each link with a plain HTTPS GET; it redirects to the file. Save the main file as file_name and each extra file at its path relative to it (textures/wood.png goes in a textures folder beside it), then import the main file.
morphic_3d_list_models
ChatGPTList 3D models: your organization's own and Morphic's shared library, newest version of each. TAKES IN: optional name_contains, source (org or library), limit and page. PRODUCES: each model's model_ref, name, source, category and height in metres, plus next_page; it changes nothing. A model_ref places the model in a 3D scene (an object's assetId) and downloads it with morphic_3d_get_model_download.
morphic_3d_list_scenes
ChatGPTList 3D scenes, most recently edited first. TAKES IN: optional project_id and limit. PRODUCES: each scene's scene_id, name and project_id; it changes nothing.
morphic_3d_prepare_model_upload
ChatGPTPrepare a private upload of one 3D model and its preview image, for example a selection exported from Blender. TAKES IN: model_size_bytes of a self-contained glTF binary (.glb, textures embedded as PNG or JPEG), preview_content_type and preview_size_bytes of a PNG, JPEG or WebP preview. PRODUCES: upload_id and two expiring PUT targets (model_upload, preview_upload); nothing is published or charged yet. PUT the .glb bytes to model_upload.url and the preview bytes to preview_upload.url, each with EVERY header of its target and no Morphic authorization header; the sizes must match exactly. Then call morphic_3d_publish_model with upload_id. From Blender, export with export_scene.gltf in GLB format, +Y up, and render a small preview (about 512 pixels) from a camera framing the model. Never pass file bytes or base64 through tool arguments. URLs are temporary credentials: do not display or log them.
morphic_3d_publish_model
ChatGPTPublish an uploaded 3D model to the organization's library. TAKES IN: upload_id from morphic_3d_prepare_model_upload after both PUTs succeeded, a name, optional project_id, and optional category, motion and pivot. PRODUCES: exactly one new 3D model and its model_ref; its animations become clips scenes can play. Charges no credits. The server reads the model itself, so a file it cannot read (missing textures, an unsupported extension) is refused with the reason; export again and upload a new upload_id. An upload publishes once: publishing the same upload_id again, or while its first publish runs, is refused.
morphic_3d_read_scene
ChatGPTRead one whole 3D scene, for example to rebuild it in Blender or before replacing it. TAKES IN: scene_id. PRODUCES: the scene (settings, objects with transforms, lights, models and keyframes, cameras with cuts, groups) and base_hash; it changes nothing. Units: metres, +Y up, +Z forward, rotations in degrees, times in seconds. Blender is Z up: Blender (x, y, z) is scene (x, z, -y). An object shows a 3D model when its assetId is a model_ref. Pass base_hash to morphic_3d_replace_scene.
morphic_3d_replace_scene
ChatGPTReplace one 3D scene's whole content. TAKES IN: scene_id, the base_hash from the morphic_3d_read_scene the new scene was built from, and the whole scene in that shape. PRODUCES: the replaced scene's object_count and its new base_hash; an open editor shows the change. Charges no credits. Everything the new scene omits is removed, so start from the read and change only what you mean to. If the scene changed since that read, nothing is written: read it again and reapply your change. A scene the engine refuses is not written, and the refusal names the field.
move_assets
ChatGPTFile library assets into a folder of their own project, or take them out of any folder. The library view makes this call. TAKES IN: up to 10 asset_ids and a folder_id from list_project_folders, or null. PRODUCES: the ids and the folder; when any asset is outside the folder's project, nothing moves.
open_library
ChatGPTOpen the connected organization's Morphic library as a full-screen browser of its images, video, audio and documents, newest first. TAKES IN: nothing. PRODUCES: the library view with its first page of assets, on clients that draw MCP App views; you receive only how many assets it shows and whether more exist, never the rows. Where no view is drawn, it shows nothing. Call it when the user asks to open or browse their library. When you need the assets yourself, to answer about them or to use one, call list_assets instead.
outpaint_image
ChatGPTExtend an image past its own edges: reframe it to another shape, or add room on the sides you name. TAKES IN: exactly one image asset id, plus either an aspect_ratio to reframe to or expand_by_pixels naming how much to add on each side, and optionally what belongs in the new room. PRODUCES: exactly ONE new image per call, the original untouched inside it and the new area filled by the model; the source is unchanged; the result is a generation_id for optional status checks with get_generation_status. Outpainting only adds: it never crops, so an image already in the asked shape is refused. To change what is inside the picture instead, use edit_image.
plan_media
ChatGPTCheck a proposed media workflow before spending credits. TAKES IN: up to 20 named steps, each with a public tool name, its arguments as a JSON object string, and dependencies. PRODUCES: execution order and, per step, validation, required scopes, the model a generation would use, a credit estimate and that step's gaps. No generation starts, files are created, or credits are charged. It does not execute the plan or seek a separate model's advice. Local edits and compositions, video inspection of existing assets, text-only images, and video generations, edits and draft finals are checked and priced as execution would; a video settled against its produced length, or whose source execution measures first, is conditional. Other source-dependent generation, voice options, audio probing and future outputs may have only conditional or unavailable estimates. complete_estimate: false means the reported sum is NOT the total budget. A dependent step may omit only the asset IDs its dependencies' outputs will fill; re-plan with the actual IDs before starting it. Client approval rules still apply to each execution tool; planning grants no permission and reserves no price.
prepare_media_upload
ChatGPTPrepare a private temporary upload for an attached or local image, video, audio or document (.pdf, .txt, .md, .srt) file. TAKES IN: filename, content_type, size_bytes and optional project_id. PRODUCES: expiring PUT credentials and an import_url; no library asset or generation yet. Spends a rate-limit slot, not credits. When the host hands an attachment over as a file object (download_url, file_id), pass it as import_media_from_url's file instead. Otherwise use only when your client can read the file's actual bytes and make an HTTPS PUT. Upload those bytes to upload_url with EVERY returned header, without Morphic authorization headers. After HTTP success, call import_media_from_url with import_url, name and project_id from this result. That call validates the bytes and returns the asset ID for edit, animate, transcribe or other media tools. An attachment or local path is not itself a transferable URL. Never substitute a similar library asset for it. Never invent bytes or pass base64 through tool arguments. If the client cannot transfer the file, ask for a Studio upload or accessible HTTPS file URL. URLs are temporary credentials: do not display or log them. If the PUT or the import fails, nothing was stored: retry that one step with the same URL and headers, copied exactly, and prepare and upload again only after its upload_expires_at or import_expires_at. After an import response you did not receive, use list_assets before retrying.
rename_asset
ChatGPTRename one library asset, which marks it modified. The library view makes this call. TAKES IN: asset_id and the new name. PRODUCES: the asset's row as browse_library returns it.
restore_assets
ChatGPTBring back assets deleted from the library in the last 90 days, with their versions but not their shares or timeline clips. The library view makes this call to undo its own delete. TAKES IN: up to 10 asset_ids. PRODUCES: restored_ids, and each other id with the reason it stayed as it was.
separate_stems
ChatGPTSplit a mixed track into its stems: the vocals and the instrumental, or six stems. TAKES IN: exactly one audio_asset_id (the mixed track); optionally a model_name from list_models under STEM, each model one split whose best_for names its stems. PRODUCES: ONE generation per call whose status lists every stem as an asset, and by name and link under stems; the track is unchanged; the result is a generation_id, and a status read with get_generation_status is how you reach the stems. For a video's sound, run extract_audio first. A stem whose asset_id is null reaches another tool only through import_media_from_url on its url. The charge is worked out from the track's length at the start, so a track longer than the model takes is refused before anything is spent.
show_asset
ChatGPTDisplay a stored Morphic asset in a media card, including an imported final edit. TAKES IN: one asset_id. PRODUCES: its media, or a document's download link, with metadata; never a document's contents. Sources are unchanged and no credits are used. Use when the user wants to see an asset that has no generation_id, such as an imported video. A new generation already owns a card, so it needs no second one. Clients without widgets receive the same asset and usable file link.
show_generation
ChatGPTShow an earlier generation's result again, only when the user explicitly asks to see it. TAKES IN: generation_id from a generation tool or list_generations. PRODUCES: status and assets, drawn as a card on a client that shows cards, read-only. Use wait_for_generations or get_generation_status for completion checks, never this.
tag_assets
ChatGPTAdd and remove the organization's tags on library assets in one step. The library view makes this call. TAKES IN: up to 10 asset_ids, and up to 10 add_tag_ids and as many remove_tag_ids from list_asset_tags. PRODUCES: each asset's tags the caller can see afterwards; a call naming any asset or tag it cannot use changes nothing.
transcribe_media
ChatGPTWrite out what is said in a recording or a video, as a subtitle file. TAKES IN: exactly one media_asset_id, an audio or a video asset from the library; optionally a model_name from list_models under TR. PRODUCES: exactly ONE new document asset per call, an SRT subtitle file with its timings; the source is unchanged; the result is a generation_id for optional status checks with get_generation_status, whose URL downloads the file. Every model reads audio and video alike, so the kind given does not narrow the choice. The charge is worked out from the source's length.
upscale_image
ChatGPTEnlarge one of the organization's images by a factor with an upscale model, and place the result in a Morphic project. TAKES IN: image_asset_id (an image asset id from list_assets or a finished generation), plus optional factor (2, 3 or 4; each model offers a subset and has its own default), model_name (from list_models under the UPS feature; only models that take an image apply) and project_id. PRODUCES: exactly ONE new image asset per call, sized at the source dimensions times the factor, started asynchronously; the source is unchanged. The result carries the generation_id for optional status checks with get_generation_status and the credits the start will charge. Use upscale_video for a video. To repair an image at its own size instead — denoise, deblur, restore — use enhance_image. Do not call upscale_image again for the same request.
upscale_video
ChatGPTEnlarge one of the organization's videos by a factor with a video upscale model, optionally interpolating to a higher frame rate, and place the result in a Morphic project. TAKES IN: video_asset_id (a video asset id from list_assets or a finished generation), plus optional factor (2, 3 or 4; each model offers a subset and has its own default), output_fps (a frame rate the model offers; omit to keep the source's), model_name (from list_models under the UPS feature; only models that take a video apply) and project_id. PRODUCES: exactly ONE new video asset per call, sized at the source dimensions times the factor, started asynchronously; the source is unchanged. The result carries the generation_id for optional status checks with get_generation_status and the credits the start will charge. Video upscaling is priced per second of the source on most models and as a flat charge on some; list_models carries each one's price. It can take several minutes; do not call upscale_video again for the same request. Use upscale_image for an image. To repair a video at its own size instead, use enhance_video.
video_change_frame_rate
ChatGPTChange one video's frame rate, dropping or repeating frames; the length and the picture are unchanged. TAKES IN: one video asset UUID and fps. PRODUCES: one NEW video; the source is unchanged. Holds the FilmLab render minimum, then settles credits to actual render compute. Returns generation_ids.
video_change_speed
ChatGPTSpeed up or slow down one video; its length changes with the speed (2 plays twice as fast and halves it). TAKES IN: one video asset UUID and factor (0.25 to 4). PRODUCES: one NEW video; the source is unchanged. Holds the FilmLab render minimum, then settles credits to actual render compute. Returns generation_ids.
video_compose
ChatGPTJoin, repeat or trim existing video clips into one video, with optional audio, transitions and overlays: each timeline clip plays its whole source or only a range of it. TAKES IN: a timeline with exactly one video track and at most one audio track, whose clips name source asset UUIDs and, to play part of one, its clip_start_ms and clip_end_ms. PRODUCES: one NEW video; sources are unchanged. All clip and image-overlay IDs are asset UUIDs. Clips on a track must not overlap; transitions need source handles around the cut. Omit resolution and aspect_ratio for ordinary joins or repeats: clips of one size keep it, and mixed sizes take the first clip's shape at its closest tier. fps defaults to 30. Gaps, trims and audio use milliseconds. To play part of a source, or to cut one clip on its own, set clip_start_ms and clip_end_ms on its clip; to repeat that part, add the same asset_id and range again at the next start_ms. Use the supplied asset IDs and cut times directly: composition reads each source's size and trims it itself. It is billed by the video it renders, not its clip count, so repeating a trimmed range costs nothing extra, while still frames or a transformed copy made first are extra paid generations; estimated_credits is the render's minimum, which an HEVC source raises. Seconds 2-4 of A twice: clips [{"asset_id":"A","start_ms":0,"duration_ms":2000,"clip_start_ms":2000,"clip_end_ms":4000},{"asset_id":"A","start_ms":2000,"duration_ms":2000,"clip_start_ms":2000,"clip_end_ms":4000}].
video_flip
ChatGPTMirror one video left to right or top to bottom. TAKES IN: one video asset UUID and axis. PRODUCES: one NEW video; the source is unchanged. Holds the FilmLab render minimum, then settles credits to actual render compute. Returns generation_ids.
video_frame_to_image
ChatGPTSave chosen frames of a video as NEW still images. Never a step toward joining, repeating or cutting a video: video_compose cuts its clips itself. TAKES IN: one video asset UUID and first, last, timestamp_ms or frame_number selection. PRODUCES: one NEW image per selected frame; the video is unchanged. For multiple timestamps, use timestamps_ms in one call: one generation_id, 1 credit per 4 frames (min 1), one asset per frame; PNG up to 5 frames, JPEG above. Never save a frame only to look at or check a video before editing it: that is a paid generation, and every operation reads its sources itself.
video_remove_audio
ChatGPTRemove the sound from one video, keeping the picture. TAKES IN: one video asset UUID. PRODUCES: one NEW silent video, or none when the video already has no audio track; the source is unchanged. Holds the FilmLab render minimum, then settles credits to actual render compute. Returns generation_ids. A video with no audio track starts nothing and is not charged: the result is outcome "unchanged" with the source as asset.
video_resize
ChatGPTResize one video without generating new content: the picture fits inside the size you give, or is stretched to it when maintain_aspect is false. A resize cannot reframe: to make a video vertical, square or another shape, use edit_video with aspect_ratio. TAKES IN: one video asset UUID and a width, a height or both. PRODUCES: one NEW video; the source is unchanged. Holds the FilmLab render minimum, then settles credits to actual render compute. Returns generation_ids.
video_reverse
ChatGPTPlay one video backwards, sound included. TAKES IN: one video asset UUID. PRODUCES: one NEW video; the source is unchanged. Holds the FilmLab render minimum, then settles credits to actual render compute. Returns generation_ids.
video_rotate
ChatGPTRotate one video clockwise by a quarter, half or three-quarter turn. TAKES IN: one video asset UUID and degrees (90, 180 or 270). PRODUCES: one NEW video; the source is unchanged. Holds the FilmLab render minimum, then settles credits to actual render compute. Returns generation_ids.
video_to_alpha
ChatGPTKey a video onto a transparent background: the people in it, what they sit on, and any prop named, over alpha. TAKES IN: exactly one video_asset_id from the library; optionally keep, the nouns to leave in the picture. PRODUCES: exactly ONE new video asset per call, a ProRes 4444 .mov with a straight alpha channel the same size as the source, with a grey matte .mov and a VP9 .webm with alpha filed beside it; the source is unchanged; the result is a generation_id for optional status checks with get_generation_status. The matte keeps only what is named: person and chair by default, so name every prop a person handles ("radio", "guitar") in keep, and never a background. It works on any backing, best on a green or blue screen, and never on a black-and-white mask video. The charge is worked out from the video's length and size at the start, and a video longer than the model takes, or too large to render, is refused before anything is spent.
wait_for_generations
ChatGPTHold the call until several generations finish, then return each one's state and assets. TAKES IN: generation_ids (1 to 10, from any generation tool) and optional wait_seconds (0 to 45, default 45). PRODUCES: one state per generation in the given order (status, error_message and finished assets with public URLs) and all_finished, read-only. Call it when the server instructions say to follow a start, the user asks for status or a link, or a requested next step needs the finished output; never just to display a result a card already shows. Prefer one call here over a get_generation_status call per generation. When all_finished is false after the wait, call again with the unfinished ids.