Summary

Built for a chat window, not an API script.

Parallel batches, editing images you attached to the message, and a history grid you can branch variations from are all shaped around a conversation rather than a request/response loop. Video is the part to plan around: renders take minutes and long clips up to about an hour, so the tool returns a pending `request_id` and the result survives closing the chat. Batch cheap drafts first, then enhance the one you want.

What it is

The hosted FLUX MCP server from Black Forest Labs. It exposes the FLUX toolkit — text-to-image, editing, multi-reference composition, virtual try-on, video and history — over a remote OAuth connection, so there is no API code and no key pasted into the conversation.

What you get

1–8 images generated in parallel from one prompt, covering text-to-image, edits, multi-reference composition, style transfer, inpainting-style edits and outpainting — `generate_image`
Virtual try-on that dresses a person or pet in a garment, hat, sunglasses, shoes or bag from a reference image, preserving face, hair and pose — `vto`
1–4 videos in parallel with FLUX 3 — text-to-video, image-to-video from keyframes, or continuation of an existing clip, 5–20 seconds with synchronized audio — `generate_video`
A draft video re-rendered at full quality from the same composition, seed and prompt plan — `enhance_video`
More images in the same direction as an earlier result, identified by `request_id`, defaulting to 4 and capped at 8 — `generate_variations`
Past generations browsed as a thumbnail grid with per-tile actions, keyset pagination and date filters — `get_history`
Your remaining BFL credit balance — `get_credits`
Model choice on `generate_image`: `flux2_pro_preview` by default, `flux2_max` for final assets, `flux2_klein_9b_preview` and `flux2_klein_4b` for speed, `flux2_flex` for readable typography; the full catalog is also the `bfl://models` resource

Requirements

A BFL account, authorized over OAuth, and a BFL organization to bill — you pay Black Forest Labs directly, with no middleman and no shared quota. Connect to `https://mcp.bfl.ai` over streamable HTTP: in Claude it is Settings → Connectors → Add custom connector; in Claude Code, `claude mcp add --transport http FLUX https://mcp.bfl.ai`; in Codex, `codex mcp add flux --url https://mcp.bfl.ai` then `codex mcp login flux`. Clients that only do stdio can bridge with `npx -y mcp-remote https://mcp.bfl.ai`, which caches tokens to `~/.mcp-auth/`. Server version 1.2.0, MIT licensed.

Setup effort

One command — claude mcp add --transport http FLUX https://mcp.bfl.ai