LoginSign Up
← Back to all sparks
C

ComfyUI

DESIGN
Velocity7.5

Node-based stable-diffusion UI for building custom image-generation workflows visually.

www.comfy.org ↗

ComfyUI is turning its node graph into something agents drive, and just open-sourced the piece that lets them do it locally.

generative-aiworkflow-automationmcpopen-weightslocal-inferencevideo-generation
◆Current state
ComfyUI absorbs new open and closed generative models continuously — LTX-2.5, Wan Animate 2, MiniMax H3 and MiniMax Music 3 as day-0 or open-weights arrivals, Seedance 2.5 and FLUX 3 as Partner Nodes — with the recurring pitch being that open weights run on hardware you own. The newest post extends that argument to the agent surface: Comfy MCP is now open source and runs against a local ComfyUI, so an agent builds workflows around the user's own GPU, models, and custom nodes from Claude, Cursor, Codex, or any MCP client.
◆Where it's heading
Three lanes are now visible and they are converging. The model lane keeps widening — MiniMax Music 3 brought song generation into a catalog that had only ever treated audio as something attached to video. The agent lane began with a hosted Comfy MCP and has now moved local and open source, which is the more consequential form: the same reasons users prefer open weights on their own hardware apply to the agent that orchestrates them. The third lane is institutional — a teams product and a university partnership — and it suggests the project is building for organizations that need the local-execution story rather than only for individual creators.
◆Prediction
With MCP running locally against user-owned models, the natural next step is agent access to the custom-node ecosystem itself rather than just workflow assembly; the Partner Node roster should keep absorbing closed models on announcement day in parallel.

◆Recent moves

  1. 1d ago

    Open Sourcing Comfy MCP on Local

    Comfy MCP is open-sourced and pointed at a local install, so an agent in Claude, Cursor, or Codex builds workflows against the user's own GPU, models, and custom nodes. It extends the MCP integration this feed already covered in June rather than opening the agent lane, but moving it local and open source is the version that matches the project's open-weights pitch.

    View source ↗
  2. 4d ago

    The 8 Best AI Creative Workflow Platforms in 2026

    A ranked listicle of competing creative workflow platforms, published on the same feed as the product announcements. Marketing content with no product change behind it.

    View source ↗
  3. 6d ago

    MiniMax Music 3: State of the Art Open Weight Music Generation

    ⚡ SPARK

    The first model in the catalog whose output is a song rather than a picture or a clip. Audio had appeared before only as a track riding along with a video model, so this opens a modality rather than extending one.

    View source ↗
  4. 7d ago

    LTX-2.5 Day-0 Support in ComfyUI

    LTX's open video model arrives on release day with multi-shot generation and a checkpoint built for fine-tuning. Another video model in a catalog that already has several, but the fine-tuning-ready checkpoint is the detail that matters for users building on their own data.

    View source ↗
  5. 11d ago

    Wan Animate 2 is now available in ComfyUI

    A character animation model that consumes driving videos directly, dropping the intermediate motion extractors that pipelines normally need. It is another video node in catalog terms, though removing a preprocessing stage simplifies the graphs users have to build around it.

    View source ↗
  6. 11d ago

    Seedance 2.5 is now available via Partner Nodes

    Seedance 2.5 joins as a Partner Node with 30-second single-shot generation and up to 50 reference assets. Partner Nodes are the closed-model half of the catalog, and this one landing hours before an open-weights post is the contrast the feed keeps drawing.

    View source ↗