Hf Studio

What hf-studio does

The hf-studio MCP server provides an agent-facing interface for creating video, images, and audio through the Higgsfield and ElevenLabs APIs. It runs as part of a self-hosted application that also includes a REST API and web UI, with generated files collected in one local library.

The project covers Higgsfield's catalog of video, image, and custom-reference endpoints, including text-to-video, image-to-video, frame-based generation, video editing, motion transfer, object replacement, text-to-image, and image editing. ElevenLabs adds text-to-speech, sound effects, music, voice isolation, and voice changes for video segments. Local ffmpeg tools support extracting frames, combining videos, and mixing audio into video.

This fits the Multimedia Process category for workflows that combine agent requests with media generation, review, and follow-up processing.

How it works

Paid operations use a quote-first flow. An agent must obtain a single-use quote_id for the exact request before starting the generation, and the server instructs the agent to show the cost and wait for approval. The estimate_cost capability reports available provider options, prices, savings against Higgsfield, and reasons a provider is unavailable.

Jobs are asynchronous and persisted in a database. Validation happens before spending, while deduplication and Idempotency-Key handling help prevent duplicate submissions. A local queue follows the account's concurrency limit, and output files are copied locally. Higgsfield jobs continue being tracked after a restart; audio jobs interrupted by a restart are marked failed rather than resumed.

Optional APIMart and KIE credentials let the hf-studio MCP server compare equivalent models and settings across providers. It can choose a lower-priced option, subject to approval rules, and can fall back after a provider rejection or an uncharged task failure. Requests that may already have reached a provider are not retried automatically.

Setup and configuration

Higgsfield access requires HF_API_KEY. APIMART_API_KEY and KIE_API_KEY are optional provider credentials. The application runs on Windows, macOS, and Linux. Provider credentials can be added with the CLI or through the web UI's Providers page; the CLI can check a key and display provider status and balances.

An optional ANTHROPIC_API_KEY enables the assistant node, which can help write or improve prompts, describe images, and draft scripts. The README identifies Claude Code, Codex, Claude Desktop, and ChatGPT as agent clients for the MCP interface.

Tools and capabilities

The hf-studio MCP server supports workflows built around:

• Video and image generation across the Higgsfield catalog.
• ElevenLabs speech, music, sound-effect, and voice features.
• Cost estimation and approval before paid execution.
• Kling 3.0 elements managed with create_element.
• Persistent jobs, local output copies, and provider-aware fallback.
• Multi-step Spaces, batch lists, reusable Flows, and built-in media presets through the web UI.

Limitations and notes

Provider availability depends on API keys, account balance, model equivalence, and request settings. A provider is excluded when it cannot reproduce the requested settings, such as a seed, bitrate, or aspect ratio. Prices may change between quotation and submission; the server checks them again before sending. Some providers may place an upfront hold, which requires separate approval.

Higgsfield retains jobs for seven days, so local copies matter for longer-term access. Higgsfield's public API cannot create or read Kling 3.0 elements, so element-based requests use APIMart or KIE. The project is MIT-licensed and requires Python 3.12 or newer.