How to Use Higgsfield with Claude
Connect Higgsfield's image and video models to Claude via MCP and direct cinematic generations from a chat.
What Higgsfield actually is
Higgsfield is a generative media platform built around 30+ underlying image and video models, packaged into purpose-built tools rather than a single generic prompt box. The headline tools are Cinema Studio (camera moves, lighting, and shot composition modelled on real cinematography techniques), SOUL ID (a consistent, reusable AI character/persona across generations), Higgsfield Audio (sound and voice generation to pair with video), and a Sora 2 Enhancer (post-processing pass that improves coherence and fidelity on Sora-generated clips). The platform's pitch is that good-looking AI video needs cinematographic control, not just a better diffusion model.
Connecting Higgsfield to Claude via MCP
Higgsfield exposes a hosted MCP (Model Context Protocol) server at https://mcp.higgsfield.ai/mcp, which means Claude can call its generation tools directly inside a conversation instead of you switching to a separate app. Add it once from the terminal: `claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcp`. The `--scope user` flag makes the connection available across all your projects, not just the current directory. This works the same way whether you're using Claude.ai web chat, Claude Cowork, or Claude Code.
Authentication: no API key to manage
Unlike most MCP integrations that require you to generate and paste an API key, Higgsfield's server authenticates through a browser-based OAuth flow. The first time Claude tries to call a Higgsfield tool, it opens a login prompt in your browser; once you approve it, the session is stored and reused for future calls. There's no key to copy, rotate, or accidentally commit to a repo — authentication state lives with your Higgsfield account, not in a config file.
Generating with a conversational prompt
Once connected, you describe what you want in plain language and Claude routes the request to the right Higgsfield tool — you don't need to know which of the 30+ underlying models to pick. For a single image: "Generate a SOUL ID portrait of a calm narrator character in soft daylight, then give me three variations with different camera angles." For video: "Use Cinema Studio to create a 10-second tracking shot moving through a neon-lit street market, dusk lighting, slow dolly-in." For audio: "Add an ambient ‘rainy café' soundscape to that last clip." Claude handles the back-and-forth — refining a prompt, retrying a failed generation, or chaining a video output into the audio tool — inside the same conversation.
Resolution, duration, and platform limits
Higgsfield's current generation limits: images up to 4K resolution, video clips up to 15 seconds. These caps apply regardless of which underlying model handles the request, so plan multi-shot sequences as separate 15-second generations stitched together afterward rather than expecting one continuous long-form clip. SOUL ID consistency holds best within a single session — for a character that needs to recur across many separate generations over time, save and reuse the SOUL ID reference rather than re-describing the character from scratch each time.
Production tips and where this fits
Use Higgsfield through Claude for storyboarding, ad creative drafts, social video concepts, and rapid client-facing previews — places where iteration speed matters more than frame-perfect final output. For final deliverables, treat AI-generated video the way you'd treat a rough cut: review for continuity errors (a known weak spot for all current video models across longer sequences), check audio sync if you layered in Higgsfield Audio separately, and keep a human edit pass before anything ships externally. Because the MCP connection is account-scoped, usage and any billing happen through your Higgsfield account, not through Claude — check Higgsfield's own pricing page for current generation costs.
More Intermediate guides
Build Your First AI Agent
From a single LLM call to an autonomous multi-step agent — the practical path.
RAG Architecture: Ground AI in Your Data
Build retrieval-augmented generation pipelines that give models access to your documents.
Fine-Tuning vs. Prompting
When a well-crafted prompt is enough — and when you genuinely need a custom-trained model.