FLUX

The FLUX MCP server exposes the full FLUX toolkit — text-to-image, image editing, multi-reference composition, virtual try-on, FLUX 3 video, variations, and history — to any client that speaks the Model Context Protocol. Generate options in parallel, edit attached images through prompts, try garments on a photo, animate stills into clips with audio, and branch into variations from any result you like. No API code, no keys pasted into the conversation.

• Hosted, remote, OAuth-only. Connect to https://mcp.bfl.ai. Sign in with your BFL account, pick the org to bill — done.
• Every FLUX.2 model. Pro, Max, Klein (4B & 9B), Flex. Your client picks the right one for the task.
• FLUX 3 video. Text-to-video, image-to-video from keyframes, and video continuation with synchronized audio, draft previews included.
• Built for chat. Up to 8 images in parallel per prompt. Edit attached images. Browse and reuse history.
• Specialist prompting guides. Official BFL guides cover focused image and FLUX 3 workflows.
• You pay BFL directly. No middleman, no shared quotas. Current rates: bfl.ai/pricing.

Install

1. Open Settings → Connectors.
2. Click Add custom connector.
3. Name it FLUX, URL https://mcp.bfl.ai.
4. Click Connect, sign in, choose the BFL organization to bill.

bashCopy code
claude mcp add --transport http FLUX https://mcp.bfl.ai

A browser opens on first use for OAuth. Tokens refresh automatically.

One-click: Add to Cursor — opens Cursor with the FLUX server pre-filled.

Manual — add to .cursor/mcp.json:

jsonCopy code
{
  "mcpServers": {
    "FLUX": {
      "url": "https://mcp.bfl.ai"
    }
  }
}

bashCopy code
codex mcp add flux --url https://mcp.bfl.ai
codex mcp login flux

Or append to ~/.codex/config.toml:

tomlCopy code
[mcp_servers.flux]
url = "https://mcp.bfl.ai"

Edit ~/.codeium/windsurf/mcp_config.json (note Windsurf uses serverUrl, not url):

jsonCopy code
{
  "mcpServers": {
    "FLUX": {
      "serverUrl": "https://mcp.bfl.ai"
    }
  }
}

Reload plugins from the Cascade panel. The first FLUX call opens a browser for sign-in.

For clients that support stdio MCP but don't handle the OAuth flow themselves — for example Hermes (Nous Research) and similar tools that only accept static bearer tokens:

jsonCopy code
{
  "mcpServers": {
    "FLUX": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.bfl.ai"]
    }
  }
}

mcp-remote opens a browser for OAuth, caches tokens to ~/.mcp-auth/, and refreshes them automatically.

Showcase

Generate a batch of options in one prompt

"Generate 4 editorial portrait variants — varied lighting, palette, and mood."

Or just one — when a single hero shot is what you need

"A top-down aerial photograph of shallow desert salt pools, intense direct sunlight."

Edit any image directly in chat

"Add a camel on the sand."

Browse and reuse your full generation history

"Show me my recent FLUX generations."

Every generation is saved. Tap any tile to edit, vary, or download. Variations branch off any past image in one click.

Tools

Your client picks which tool to call based on your prompt — you don't need to invoke them by name.

Models

Available on generate_image:

The full catalog, per-model reference-image limits, the FLUX Tools (vto), and FLUX 3 video are also exposed as the bfl://models MCP resource.

Video

generate_video renders 24fps MP4 clips of 5–20 seconds with synchronized audio, up to 4 in parallel per call. Every request uses one of three modes:

• Prompts are prose FLUX 3 reasons about the prompt and plans the generation before rendering — describe the scene, the motion, and the sound. There is no seed or model parameter on video.
• Keyframes reference a prior result (request_id), an uploaded image, or a public URL. Optional per-keyframe timestamps (in seconds, ascending — set on all keyframes or none) pin exact positions in the clip; with duration set to auto, the clip runs to the last timestamp (20 seconds max).
• Source videos for v2v are a prior generate_video result, an attached MP4 up to 50 MB, or a public MP4 URL.
• Format controls. duration: 5–20 seconds or auto (default). aspect_ratio: auto (default), 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, or 9:16. resolution: hd (default, 704p-class) or fhd (finished by a video upsampler on a 1080p-class canvas). generate_audio: on by default.
• Drafts are the way to explore. draft: true renders a fast low-step preview of the exact planned generation (hd only). Batch cheap drafts, pick the winner, and enhance_video re-renders it at full quality from its request_id.
• Renders take minutes — long clips up to about an hour. The tool returns immediately with a pending request_id; results survive closing the chat and resume automatically.
• Reference-based conditioning (identity and motion references) is not part of the video surface; a dedicated r2v mode follows later.

Prompting skills

The server ships specialist prompting guides from black-forest-labs/skills and serves them three ways, so every client gets at least one. These guides cover focused workflows; a general FLUX 3 prompting guide is not included yet:

• Tools — list_skills and get_skill(name). Works everywhere, including ChatGPT.
• Resources — bfl://skills/<name> for clients with resource pickers (Claude, Cursor).
• Prompts — skill_<name_with_underscores> slash-prompts where the client supports them (e.g. skill_flux_image_best_practices).

Ask "what FLUX skills do you have?", or just request something specialized like a product ad, Super 8 footage, dialogue direction. The generation tools point your client at the matching guide before it writes the prompt.

Example prompts

Try these in your client once FLUX is connected:

Copy code
Generate 4 editorial portrait variants — varied lighting, palette, and mood.
Use FLUX.2 [pro] at 1440×1792.

Copy code
A top-down aerial photograph of shallow desert salt pools at golden hour,
intense direct sunlight, #FF6B6B and #F5E6D3 mineral deposits, FLUX.2 [max].

Copy code
[attach an image]
Add a camel on the sand. Preserve the rest of the composition exactly.

Copy code
Generate a movie poster with the title "MIRAGE" in 1970s grindhouse type,
high-contrast cyan and orange. Use FLUX.2 [flex].

Copy code
[attach a selfie]
Try this on me — a red Nike Tech Fleece hoodie, oversized fit.

Copy code
[attach an image]
Animate this into a 10-second clip — slow push-in, wind through the palms,
distant surf.

Copy code
Generate 3 draft videos of a paper lantern rising over a night market,
each with a different camera move. I'll enhance the best one.

Copy code
Show me my recent FLUX generations.

Copy code
[click Variations on a past image]
4 more like this one, vary the lighting.

Prompt tips

• Front-load the subject. Put the most important object, person, or scene first.
• Describe lighting. "Soft golden hour light" or "overcast diffused studio light" gives the model useful direction.
• Use hex colors. #FF6B6B (coral pink) is more precise than "pinkish red".
• Quote rendered text. Use exact quoted strings for typography, labels, posters, and signs.
• Avoid negative prompts. FLUX responds to what you describe, not a list of what to avoid.
• Iterate from results. Use Variations for alternatives, or Edit to keep refining a generated image.

The server exposes specialist guides for focused workflows. A general FLUX 3 prompting guide is not included yet.

Troubleshooting

• Claude Desktop / claude.ai: Settings → Connectors, confirm FLUX shows as Connected. If it failed silently, remove and re-add — make sure pop-ups aren't blocked so the OAuth window can open.
• Claude Code: claude mcp list should show the server.
• Codex: codex mcp list should show flux. Start a new Codex session after adding.
• Quick smoke test: ask your client "check my BFL credits".

• Claude.ai / Desktop: Settings → Connectors → toggle FLUX off and on, or click Reconnect.
• Claude Code: /mcp shows status and reauth. Full rebuild:

  bashCopy code
  claude mcp remove FLUX
  claude mcp add --transport http FLUX https://mcp.bfl.ai
  

• Codex: codex mcp login flux. Full rebuild:

  bashCopy code
  codex mcp remove flux
  codex mcp add flux --url https://mcp.bfl.ai
  

• mcp-remote clients: rm -rf ~/.mcp-auth and restart the client to force a fresh OAuth.

• Make sure you have a BFL account at bfl.ai.
• Disconnect and reconnect the connector to redo OAuth.
• Check the selected organization has sufficient credits.
• Ask your client to check your BFL credits to verify the balance.

Large batches, FLUX.2 [max], or complex edits take longer than smaller generations. Visual MCP clients update the image view automatically.

Video renders take minutes — long clips up to about an hour. They return a pending request immediately and resume automatically, even if you close the chat and come back later.

Your MCP client needs permission to upload attached media to BFL. If your client blocks outbound HTTPS from its sandbox, allow *.bfl.ai or pass a public URL instead. Video uploads are MP4 only, up to 50 MB.

Disconnect the FLUX connector in your client and reconnect it. The OAuth flow will prompt you to select an organization again.

Agent Skills (companion repo)

MCP and Agent Skills solve different problems:

Use both. They complement each other and the skills repo's prompting guides are served directly over MCP too (see Prompting skills).

Links

• Docs — docs.bfl.ai
• Pricing — bfl.ai/pricing
• Status — status.bfl.ai
• Black Forest Labs — bfl.ai

License

MIT © Black Forest Labs