Run Higgsfield MCP
Run Higgsfield MCP is a free agent skill maintained by Scopeful. It teaches an AI coding agent such as Claude Code, Cursor, Windsurf or Codex how to drive this tool correctly, so you do not have to re-explain it every session. Every published Scopeful skill is free and the install command is public, with no sign-in required. Install it with npx skills add higgsfield-ai/skills. Scopeful also tracks hand-verified USD pricing for 39 AI creative tools at https://www.scopeful.org/tools.
Connect any agent to the real Higgsfield MCP (HTTP at mcp.higgsfield.ai/mcp). Config snippets for Claude Code, Cursor, OpenCode, Codex, Windsurf, Gemini.
Tags: video, image, mcp, higgsfield, http
Install
npx skills add higgsfield-ai/skillsReference
name: higgsfield-mcp
description: Use this skill whenever the user wants to connect any agent (Claude Code, Cursor, OpenCode, Codex, Windsurf, Gemini) to the Higgsfield MCP server. Triggers include any mention of "Higgsfield MCP", mcp.higgsfield.ai, "add the higgsfield MCP to my editor", "Higgsfield MCP config", or asking an agent to register a remote MCP server. Do not trigger for the CLI install — that is the higgsfield-cli skill. Do not trigger for prompt structure, model selection, or Marketing Studio workflows — that is the higgsfield-mcp-better master skill.
Run the Higgsfield MCP
The real Higgsfield MCP is a remote HTTP server at https://mcp.higgsfield.ai/mcp. It is not an npx package. The cursor-plugin repo (higgsfield-ai/cursor-plugin/mcp.json) is the canonical config — verified 2026-06-30:
{
"mcpServers": {
"higgsfield": {
"type": "http",
"url": "https://mcp.higgsfield.ai/mcp",
},
},
}
The first connection opens a browser-based auth flow against the user's Higgsfield account. The token is stored by the agent and reused for subsequent calls. Sessions are short-lived; the MCP server returns 401 when the token expires and the user re-authenticates in the browser.
Verified facts (live 2026-06-30)
- MCP endpoint:
https://mcp.higgsfield.ai/mcp(HTTP, returns 401 Unauthorized without a valid bearer token). - Transport:
type: "http". Not stdio, notnpx -y @higgsfield/mcp(that npm package is not published — the existing Scopeful skill'snpx -y @higgsfield/mcpsnippet is wrong; ignore it). - Auth: browser-based OAuth on first connection, per the
cursor-pluginREADME ("The first connection prompts authentication"). - Source of truth for the config:
higgsfield-ai/cursor-plugin/mcp.json.
Assumed (not independently verified against a live session)
- The MCP wraps the same model surface as
@higgsfield/cli— 30+ models, Marketing Studio, Virality Predictor, 4 workflows. The tool names exposed by the MCP server are not enumerated in any public repo, so the only safe answer is "list the tools the MCP registers in your agent after auth" rather than committing to a fixed tool list. - Polling and result-URL behavior follow the same shape as the CLI: jobs have IDs, a
get_status(or similarly named) tool, and the terminal state includes a downloadable URL. Exact tool names need to be discovered per session. - Cost is consumed from the same Higgsfield account the CLI uses. No separate billing surface.
Install (per agent)
The config is the same for every host — only the file path changes. The mcpServers block from cursor-plugin/mcp.json is the canonical body.
Claude Code
Add to ~/.claude/mcp_servers.json (user-scope) or .mcp.json (project-scope):
{
"mcpServers": {
"higgsfield": {
"type": "http",
"url": "https://mcp.higgsfield.ai/mcp"
}
}
}
Reload Claude Code. The MCP appears in the tool list. First call opens the browser auth flow.
Cursor
Install via the Cursor Marketplace: search "Higgsfield" (/add-plugin -> Higgsfield), or copy the same JSON into ~/.cursor/mcp.json. Per the cursor-plugin README, the first connection prompts authentication.
OpenCode
Add to opencode.json at the project or user level:
{
"mcp": {
"higgsfield": {
"type": "remote",
"url": "https://mcp.higgsfield.ai/mcp"
}
}
}
Restart the OpenCode session. The MCP registers on connect.
Codex
Codex uses ~/.codex/mcp.json with the same mcpServers block. Restart after editing.
Windsurf / Gemini / others
Same JSON shape, different config-file path. The host-specific path is in each agent's docs; the mcpServers body does not change.
Verify
After saving the config and restarting the agent:
- List the registered MCPs. In Claude Code:
/mcp. In Cursor: Settings -> Tools & MCPs. In OpenCode:opencode mcp list. - The
higgsfieldentry should show a green / connected dot. - Run a tiny test through the agent: "list the Higgsfield models". The agent should call whatever
list/list_models/modelstool the MCP exposes and return the catalog. - If the dot is red, the most common causes are: (a) the agent cannot reach
mcp.higgsfield.ai(corporate proxy / firewall), (b) the auth flow was canceled (re-trigger by removing and re-adding the server), (c) the JSON is malformed (validate withjqbefore restarting).
Live feedback (what the MCP gives back)
- Job IDs come back synchronously from the create / submit tool. Save the ID.
- Use the status / poll tool every 10 seconds. Faster polling hits rate limits; the upstream Higgsfield API enforces a minimum interval.
- On
complete(or whatever the terminal-state label is in your MCP), the response includes a media URL. URLs expire (the CLI documentation says 7 days for some result types) — download or rehost immediately. - On
failed, the response includes a structured error code. Map it back to the same failure reasons documented in thehiggsfield-cliskill (NSFW filter, invalid preset, source image too small, missing required params, unknown params, etc.). - The MCP is a thin layer over the same backend as the CLI. If a failure mode is documented in the
higgsfield-cliskill under "Live feedback", the same fix applies here.
When NOT to use the MCP
- The user wants to run one-off generation commands from a terminal. Use the CLI (
higgsfield-cliskill) — it has a shorter feedback loop and the--waitflag is built in. - The user is building a server-side pipeline that already has the
@higgsfield/clientTS SDK installed. Use the SDK directly (subscribe()+withPolling: true); it is more efficient than the MCP for high-throughput jobs. - The user only needs Marketing Studio brand-kit URL fetch or one-off Virality Predictor scoring. Both have a CLI one-liner (
higgsfield marketing-studio brand-kits fetch --url ...andhiggsfield generate create brain_activity --video ... --wait) that is faster than wiring up an MCP. - The user does not have a Higgsfield account. The MCP auth flow will fail. The CLI's
higgsfield auth loginis the entry point either way.
Where to go next
- The
higgsfield-cliskill: terminal-first install + command reference for the same model surface. - The
higgsfield-mcp-bettermaster skill: model selection, prompt structure, Marketing Studio modes, Virality Predictor scoring, the polling / failure playbook, and the cross-agent install paths for thehiggsfield-ai/skillsplugin pack (higgsfield-generate,higgsfield-soul-id,higgsfield-product-photoshoot,higgsfield-marketplace-cards).