The Higgsfield MCP Connector
Thirty-plus image and video models behind one connection. Setup, the five tools, model selection, credits, and the context cost nobody mentions.
Part 1 — What It Is
Higgsfield runs a hosted Model Context Protocol server. Connect it to an MCP-capable AI client and that client can generate images, generate video, train reusable characters, and check job status directly in the conversation — without opening the Higgsfield web app.
Authentication is OAuth against your existing Higgsfield account. No API key to manage. No package to install. Your existing plan credits transfer automatically.
Which clients work
Any client that speaks the MCP standard can connect. Confirmed as of August 2026:
✓ Connects via MCP server
- Claude (web, desktop, mobile)
- Claude Code
- Cursor, Windsurf, Cline, OpenCode
- Cowork, OpenClaw, Hermes Agent, NemoClaw
⚠ Different route
- ChatGPT — connects via plugin, not MCP server
- ChatGPT desktop — did not work as of May 2026
All routes use the same Higgsfield account and credit pool.
Part 2 — Setup
Claude — web or desktop
Sign up at higgsfield.ai if you don't already have one. New accounts receive free starter credits. No credit card required to start.
Go to higgsfield.ai/mcp and copy the server URL. Current URL: https://mcp.higgsfield.ai/mcp
/mcp path. Use whichever the current official docs specify — confirm before connecting.Open Settings → Connectors. Click the + icon at the top right, then select Add Custom Connector.
In the Name field type Higgsfield. Paste the server URL into the URL field. Click Add, then Connect.
A browser OAuth window opens. Authenticate with your Higgsfield account. No API key is needed — this uses your existing account credentials.
List my available Higgsfield models and show my current credit balance.
You should get the model roster and a credit count. If you get nothing, the auth handshake didn't complete — disconnect and reconnect from Settings → Connectors.
Claude Code and terminal agents
Skip the browser flow. One command adds the connector:
The OAuth handshake happens in the browser on first use. Check current syntax in your client documentation — MCP tooling has changed as the standard has matured.
geopopos/higgsfield_ai_mcp is a community-built local MCP server that requires HIGGSFIELD_API_KEY and HIGGSFIELD_SECRET environment variables and a local Python install. That is a different product from the official hosted OAuth connector documented here.Part 3 — The Five Tools
The tool surface is small — which is a good thing. Five tools cover the entire platform. Every MCP-capable client calls the same five.
Part 4 — The Model Roster
30+ models are available through the connector. The agent selects automatically based on the task, or you can name a model explicitly in your prompt. Naming is worth doing once you know your production preferences — auto-selection is fine for exploration and less predictable for production work.
The seven most commonly used
Also in the roster
Sora 2, Wan 2.6, MiniMax Hailuo, Soul Cinema, Seedream, Flux Kontext, and 20+ additional image and video models. The full current list is at higgsfield.ai/ai-image and higgsfield.ai/ai-video.
Part 5 — The Context Cost Nobody Mentions
The Higgsfield connector loads 30+ tool definitions into your context window when connected. This is the MCP tax — it's real and it compounds on long conversations.
Context window consumption — before you type anything
Source: claudefa.st. Figures are approximate and depend on client implementation.
Practical implications
- If you run only Higgsfield, this barely matters
- If you run Higgsfield alongside several other MCP servers, enable tool search or expect a meaningful context tax on every conversation
- Disconnect connectors you aren't using in a given session
- On long conversations, the upfront token cost compounds — it's paid out of the context window you'd otherwise use for the work itself
Part 6 — Credits and Cost
Credit efficiency tips
- Use Nano Banana Pro for start-frame composition testing — lower cost, fast results
- Generate composition and blocking passes without audio enabled
- Enable audio only once the picture is close to final
- Use
get_generation_statusto poll rather than re-triggering generations - Disconnect the connector between sessions if you're running multiple MCP servers
Part 7 — What It Is Good For
Higgsfield MCP wins on aggregation and single authentication. Thirty-plus models, one connector, one credit pool, one polling pattern. The time saved on context switching between platforms is the entire ROI.
✓ Good fits
- Multi-model workflows in a single conversation
- Character-consistent campaign content via Soul ID
- Rapid iteration across image and video without tab switching
- Agentic workflows where the AI selects and chains tools
- Teams sharing one credit pool across multiple clients
✗ Not designed for
- Editing — you get generated assets, not a timeline
- Long-form video — 15 seconds is the per-generation cap
- Publishing — distribution is still your job
- Guaranteed usable output — prompting still matters
Part 8 — Two Things to Get Right Before Publishing
AI disclosure
Platform rules on labeling synthetic media are tightening. EU AI Act Article 50 applies from August 2026. Check the disclosure requirements for where the content will be published — not where it was generated. The requirement follows the destination, not the tool.
Rights and usage
Confirm Higgsfield's current terms on commercial usage and ownership of generated output before building a business on it. Terms vary between platforms and change. Read the current terms of service rather than assuming they match what you've read in a guide — including this one.
Volume without judgment
MCP connectors make generation fast enough that volume stops being the constraint. Judgment becomes the constraint. An agentic workflow that generates 50 clips unattended is only useful if someone reviews all 50 before any go out. Speed of generation does not reduce the need for editorial review — it increases it.
Part 9 — Troubleshooting and Quick Reference
| Symptom | Likely cause | Fix |
|---|---|---|
| Connected but getting no response | OAuth handshake didn't complete | Settings → Connectors → disconnect Higgsfield → reconnect → complete the browser auth window fully |
| Models listed but generations failing | Insufficient credits | Check balance at higgsfield.ai dashboard. Top up or wait for monthly reset |
| Context window filling fast | 30+ tool definitions loaded without tool search | Enable tool search in client MCP settings. Disconnect unused connectors between sessions |
| Agent keeps re-triggering generations | Not using get_generation_status to poll | Explicitly ask the agent to poll job status rather than re-generate. Video jobs take 90+ seconds |
| URL not connecting | Server URL has changed | Re-check higgsfield.ai/mcp for the current URL. Remove the connector and add fresh with the new URL |
| ChatGPT not finding the MCP server | ChatGPT uses the plugin, not MCP server | Add Higgsfield via ChatGPT's plugin directory, not via MCP connector settings |
| Wrong model being selected | Auto-selection chose a different model than expected | Name the model explicitly in your prompt: "Use Kling 3.0 for this generation" |
Quick reference
Sources
- higgsfield.ai/mcp — official MCP documentation
- Higgsfield blog — "How To Generate AI Videos Straight From Claude with Higgsfield's MCP"
- Higgsfield blog — "MCP for Marketers"
- claudefa.st — "Higgsfield MCP: Sora, Veo, Kling from Claude Code"
- TECHSY — "Higgsfield MCP for Claude Code: 60-Second Setup"
- Jonathans Blog — "How To Use Higgsfield With Claude (via MCP)"
- higgsfieldmcp.com — independent resource, not affiliated with Higgsfield
- GitHub: geopopos/higgsfield_ai_mcp — community-built local server (different product)