Higgsfield MCP server for AI agents

Search Higgsfield Marketplace apps callable through MCP. Returns each app's id, name, and the actions it exposes. Flow: apps_search to find an app → apps_describe(app_id, action) to get an action's argument schema + manifest_revision → apps_invoke to run it. Read-only; does not call any app.

Get the user's available credits and current subscription plan. For transaction history, call `transactions` instead.

Cancel the auto-renewal of the Higgsfield MCP free trial. Call this when the user asks to cancel the trial, cancel auto-renewal, stop the upcoming charge, or asks how to cancel. IMPORTANT SEMANTICS: cancelling stops the automatic charge at the end of the trial ONLY, the user KEEPS trial access and remaining trial credits until the trial ends; nothing is charged. First call WITHOUT `confirm` (or confirm=false): in UI clients this opens a confirmation card with 'Keep trial with auto-renewal' and 'Cancel auto-renewal' buttons; in text-only clients relay `assistant_response` verbatim and wait for the user's explicit confirmation. Only call again with `confirm=true` after the user explicitly confirmed the cancellation in chat. Never pass confirm=true on the first call.

Open the Create Voice Apps UI. Call this immediately when the user asks to create a voice, call the Create Voice tool, or needs a local browser record/upload surface and no confirmed audio_media_id is already present. Do not ask the user to upload an audio file or provide the name in chat first; the widget collects the required name plus record/upload audio. If the user already has or attached an audio file in chat, still call this tool with initial_tab='upload', remote tools cannot read Claude chat attachments, so the user re-selects the file in the widget's Upload tab (it uploads directly to Higgsfield). Do not try to pass a chat attachment or ask for a URL. The widget records/uploads, confirms the audio, and creates the voice itself end-to-end, and if the user is out of credits or on a free plan it shows the plans/credits UI inline. After the widget reports success you do NOT need to call create_voice_from_confirmed_audio again. Only call create_voice_from_confirmed_audio yourself when a confirmed audio_media_id is already present in the prompt and no UI step is needed.

Start a new full-stack website. Creates the website and a git repo: a React 19 + TanStack Start app, server-rendered, in ONE Cloudflare Worker, with D1 / R2 / KV / Durable Objects / Containers available (all DISABLED by default). Returns a website_id, pass it to every later website tool. The 'type' param is REQUIRED and is the USER'S choice, not yours: unless the user has already made it unambiguous, ASK the user whether they want a plain website (no Higgsfield integration) or a Higgsfield-integrated app (Sign in with Higgsfield + AI image/video generation via the Higgsfield SDK) BEFORE calling this tool. Apps are scaffolded from a v2 starter template and REQUIRE the 'template' param, pick the closest of studio / preset / app-detail per the template param's guide ('custom' is ONLY for when the user explicitly says "use custom template", never pick it yourself). The chosen layout ships as real code already wired as the home page; you ADAPT IT IN PLACE, never rebuild it. Websites take an OPTIONAL template: pass 'scroll-scrub' for an animated website (its scrub engine ships pre-built) and omit it for a non-animated one. App and website templates are not interchangeable, a cross-kind name is rejected. Workflow: (0) call get_workflow_instructions with { workflow: "website-builder-flow" } FIRST to load the stack, design contract, and hard rules (REQUIRED before building or editing); (1) create_website; (2) call website_repo_access to get the repo's git URL + scoped token, clone/edit/commit/push with the terminal (for apps, read app/src/layouts/AGENTS.md + app/src/components/AGENTS.md right after cloning); (3) deploy_website to ship it live, and deploy again after ANY later change (publish_website only lists what is already live; it does not deploy).

Generate a 3D GLB mesh. Use `models_explore(type:'3d')` to pick a model and see its `medias[].roles` and `parameters`. Apps UI local file: call `media_upload_widget`; remote tools cannot read Claude chat attachments. Web media URL: call `media_import_url`, pass returned `media_id`; `medias[].value` must be media_id/job_id, not URL. Defaults: `image_to_3d` for general image-to-3D with optional texturing, PBR, and rigging; `multi_image_to_3d` when 2-4 views of the same subject are available (better geometric accuracy); `sam_3_3d` for single-object reconstruction; `3d_rigging` to rig an existing 3D model (takes `model_url`, not images, pass a prior 3D job_id or an https GLB URL). For animated rigs, search clip ids with the `animation_actions` tool and pass `animation_action_id` with `enable_animation:true`. The mesh reproduces only what is in the source image, to add or change props, clothing, or held objects, edit the image first with `generate_image`, then convert the edited result. Pass model-specific params as top-level fields. Apply `adjustments` returned by the server. If `recovery_tool` is returned, call it immediately. `get_cost:true` preflights credits without submitting.

Generate one image request and render its result(s) in the generation widget. Use count 2-4 only for variants of the same prompt, inputs, and settings; for 2-12 independent image requests with different prompts or inputs, use the headless generate_image_batch tool instead. Apps UI local file media: call `media_upload_widget`; do not ask for Claude chat attachments because remote tools cannot read them. Web media URL: call `media_import_url`, then pass returned `media_id`; `medias[].value` must be media_id/job_id, not URL. Defaults: `marketing_studio_image` for commercial/product/ads; `soul_cast` for text-only character/avatar; `soul_2`+`soul_id` for trained reusable Soul; `soul_2`/`nano_banana_pro` for one-off character refs; `soul_2` for portraits/fashion/UGC/editorial; `nano_banana_pro` for 4K/text/diagrams. Ambiguous create-character/avatar: offer reusable Soul training (5-20 photos, ~10 min) vs one-off; do not train generic silently. Use `show_characters(action='train')` only if explicitly requested or user provides 5-20 photos. Use `models_explore` for aspect_ratios, params, medias roles. Top-level model params; apply `adjustments`. If `recovery_tool` returned, call it immediately; do not explain/ask first. `get_cost:true` preflights credits. `use_unlim` defaults false, pass true only when the user explicitly asks to use their unlimited/free-trial generations, never to save them credits on your own initiative.

Submit 1-12 independent video generations in parallel without opening a widget. Each requests[] item accepts the same params as generate_video, creates exactly one job, and keeps its caller-provided index in the response. Use for multiple distinct prompts or inputs; use generate_video for one user-facing generation. Poll returned job IDs with jobs_wait in agent-chosen groups of at most 12. For larger sets, collect indexed jobs across submission batches. After every job in the user's set is terminal, pass the collected jobs to exactly one show_generation_by_ids call for up to 60 jobs; never use show_generations or call job_display once per job.

Show the explainer video style presets (CMS-managed catalog). Returns preset ids, names, and preview media. When the user picks one, resolve it with resolve_explainer_preset to get the style reference media_id for generations.

Long-poll 1-12 generation jobs together without opening a widget. Waits up to timeout_seconds (default 15, max 15) for every job to reach a terminal state, then returns compact indexed statuses and result URLs. Use job IDs returned by generate_image_batch, generate_video_batch, or generate_audio_batch. For larger sets, choose groups of at most 12 and wait for each group. Permanent lookup failures are returned once without blocking the other jobs; transient lookup failures are retried within the timeout. When all_terminal is false, wait poll_after_seconds before calling again. After every wait group in the user's generation set is terminal, collect their indexed jobs and display them with one show_generation_by_ids call when within that tool's limit. Never use show_generations or call job_display once per batch job.

List available voices for speech and voice tools. Returns built-in preset voices plus the user's own custom voices. Each voice has a voice_id and a voice_type ('preset' or 'element'); pass that exact pair to the audio models (via generate_audio, seed_audio or text2speech_v2) and to the voice_change tool to select the speaking voice. Use the preview_url to hear a sample. Paginate with the returned next_cursor.

List every workspace the user can access (their private workspace plus any shared/team workspaces). The `is_selected` field marks which workspace MCP operations currently target. Use when the user asks which workspaces they have, or wants to switch workspace.

Upload media for use in generation, or general files (documents, archives, code) for sharing. Returns presigned URLs for clients that can upload bytes themselves; run the generated curl commands or PUT the bytes to each upload_url, then call media_confirm. The media type is inferred from the filename extension: image/video/audio extensions become generation inputs; other whitelisted extensions (pdf, zip, tar, docx, csv, code files, …) are uploaded as general files and return a permanent URL, but cannot be used as generation inputs. General files are the agent's own upload path, the widget does not accept them, so upload the bytes to upload_url yourself (e.g. from a code execution environment). Supports batch uploads via files[]. Do not use this for user-provided local image/video/audio in Claude Apps UI-capable clients; call media_upload_widget instead so the user chooses the file in the Higgsfield widget and the browser uploads it directly.

Find generation models. Use recommend with goal + input context; use get for model constraints. Items carry supports_unlim when the model accepts free-trial unlimited generations; the top-level unlim block says whether the caller can spend them right now, and the trailing 'Unlim configs' text lists the configurations their allowance actually covers.

Enter the website in the current Higgsfield app contest, together with the social-media links promoting it. A website not yet PUBLISHED to the community feed is published automatically by the entry, no need to call publish_website first. The website DOES need a live production deploy (deploy_website), else the entry is rejected. BEFORE entering, make sure the page metadata in app/src/app-meta.json is filled with real values (og_title etc.), the auto-publish lists the website on the feed and an empty og_title makes it INVISIBLE there. Pass one or more urls, each a social-media link (YouTube, X/Twitter, Instagram, or TikTok); any other host is rejected. There is a single active contest, so no contest id is needed. Calling again for the same website OVERWRITES its urls (use it to fix or add links), it does not create a second entry.

Show available Higgsfield presets for image-to-video generation. Returns preset ids, names, previews, and descriptions.

Publish the website: lists the website's CURRENT LIVE production deploy on the Higgsfield community feed ('show in feed'), where other users can discover it. This does NOT deploy, deploy_website (which every build flow already runs) must have shipped the latest changes first; publishing with undeployed changes lists the OLD live build, and re-publishing does not re-deploy. BEFORE publishing, the page metadata in app/src/app-meta.json MUST be filled with real values, og_title, og_description, favicon_url, og_image_url, the feed card renders from them (read fresh from the pushed repo at publish time) and a website with an empty og_title is INVISIBLE on the feed; the live page's own head tags are baked at build time, so deploy AFTER changing them. Also OFFER the user a cover video for the card (og_video_url), ask their permission first (video generation costs credits), never generate it unprompted. Commit and git push the metadata (and all other changes), then deploy, BEFORE calling this. Publish when the user asks to publish / share / go live on the feed, OR when they opted in to publishing at the start of the build, in that case publish automatically once the site is deployed with its metadata filled, without waiting to be asked again. For a plain deploy without a feed listing use deploy_website instead. EXCEPTION: a website whose production was never deployed (or was taken down by unpublish) falls back to deploying first, that returns status 'pending' while CI runs and the website is listed automatically once the deploy succeeds (check with website_status).

Remove or cut out the background from an existing image or video. Use this when the user asks for background removal, a transparent background, an isolated subject, a clean cutout, or a subject-only asset. Pass media_id for the source media and media_type as image or video; the matching background remover is selected automatically. This tool does not use prompt, count, or style parameters.

Confirm the user has rights to the content of an `ip_detected` generation and flip its status to `completed`. Backend accepts only seedance-family jobs (cs_3_0, seedance_2_0, ms_video, etc) and only while the job is still in `ip_detected` state. Returns the updated generation. Used by the job-list widget's Reveal button after the user accepts the rights confirmation modal.

Start a Shorts Studio short: restyle one uploaded source video (4s-120s) into a set of AI-generated short-form clips using a style preset. PAID, reserves credits. Prerequisites, gathered in whatever order fits the conversation: (1) a style preset, pick one via shorts_studio_list_presets or make one with shorts_studio_create_preset; (2) a source video, an uploaded video_input id from media_upload_widget (type=video). If the user hasn't provided a source video yet, ask them to upload one before calling this. Output orientation defaults to 9:16 (vertical); pass aspect_ratio:'16:9' for horizontal. Returns a session with empty job_ids; poll shorts_studio_status until clips appear. Set get_cost=true with duration_seconds to estimate the credit cost without submitting a job, no preset or source video needed for the estimate.

Poll one Shorts Studio session. Returns {id, status, job_ids}. status='completed' means every clip job is terminal (not necessarily successful). Poll each job_id via job_status for its clip video url and per-clip status.

Browse completed non-Marketing Studio generation history and render one paginated page in the gallery widget. Returns generations with {id, type, status, model, params, results}. Use only when the user explicitly asks to browse regular generation history. Do not use this history tool after generate_*_batch or jobs_wait; show an exact completed batch with one show_generation_by_ids call instead. Use show_marketing_studio_generations for Marketing Studio video/image/ad history. Pass a prior generation's id as value in the medias array of a new generation to reuse it. Use job_display only to inspect one specific previous result. Paginated: if next_cursor is not null, pass it as cursor to get the next page.

List your uploaded media files by type. Returns media IDs, URLs, and creation timestamps. Call once with the single type the user asked for (default image); do not enumerate the other types unless the user explicitly asks for them. Pass media IDs as value in the medias array of generation tools. Paginated: if next_cursor is not null, pass it as cursor to get the next page.

Sync Agents, imports the user's user-authored Skills and a personality dump from the current host LLM into Higgsfield. One trigger, one upload, one final confirmation. Calling modes: 1. `message: "/sync-agents"`, server returns a short ack in `content[0].text` plus an assistant-only directive containing a self-contained Python upload script (pre-wired with a one-time presigned POST URL scoped to one per-user prefix). Print the ack verbatim. Then follow the directive: enumerate user-authored skills + synthesize `personality.md` into `/tmp/profile-import/{skills,personality}/`, run the script in your code-execution sandbox, capture its single-line JSON summary, and call this tool again with that summary as `message`. 2. `message: "<script summary JSON>"`, server reads the uploaded .md files, parses each, and upserts the user's Skill rows. Returns a one-line confirmation summarising counts (e.g. "Imported 3 skill(s) (2 new, 1 updated) · personality queued for save to your memory."). Print that confirmation verbatim to the user. **Also pass `host`** on this call, your own runtime identifier (e.g. `claude_ai`, `claude_code`, `codex`, `cursor`); it is stamped on every newly imported skill's `import_origin` so the marketplace can attribute origin per host. No polling, no chains. Each call returns synchronously.

List the user's connected TikTok accounts. Returns each account's connector_id (needed by other tiktok_* tools) and status. `active` accounts are ready; `error` accounts need tiktok_reconnect; no accounts ⇒ offer tiktok_connect. Read-only.

Open the tuning editor for one Commercial Music Library track the user already picked (via tiktok_music_trending): trim start/end and set track/original volumes. Pass the same genre/country_code/date_range filters that were used when the track was found, or the lookup may miss. The user copies the final configuration from the editor and pastes it into the chat; pass those values to tiktok_publish. Read-only.

Upscale and enhance an existing image. Use this when the user asks to upscale, enhance, or increase the resolution of an image to 2K/4K. This tool does not use prompt or count. Provider selects the upscale backend; currently only 'bytedance' is supported (the default). You MUST pass the source image's width and height in pixels (the caller supplies them; the server does not infer them). Set params.get_cost=true to preflight credits (flat cost) without submitting a job.

Get the status and result of a video analysis. Poll this after video_analysis_create until status='completed' (scenes populated) or 'failed' (fail_reason populated). Analyses typically finish in 3-5 minutes, poll accordingly every 30-60 seconds.

Inspect the website's database (D1 / SQLite), READ-ONLY. The website has ONE database, the live site's real data. Pick an operation: 'tables' (list tables); 'schema' (a table's columns, needs table); 'rows' (a page of rows, needs table; optional filters, order_by + order_dir, limit (default 50) + offset); 'query' (one read-only SELECT/WITH, needs sql). Writes and DDL are rejected.