Congrats, awesome to see this ship. We clearly landed on the same idea from opposite ends: you took it to the cloud, I built it offline-first. ComfyUI - Agent Kit gives the same full natural-language control, where the agent assembles the entire workflow, generates images, video, and audio, and even builds custom nodes (V3 API) for your exact task, across Claude Code, Codex, Gemini, and Qwen. It ships near-daily and folds in community fixes and workarounds by hand, often before they reach an official build, so the agent routes around broken paths before it builds. The twist: it runs entirely on your own machine, private and NDA-safe, and still reaches online models when you want. Cloud for scale, local for control.
Hi Slava! Thank you for sharing (and contributing to the overall ComfyUI community), I just recently came across your repository and plan to dig into it over the next couple of weeks. Sounds very cool!
That automated custom node creation I mentioned is already done, v2.0.1
It came out of a real need: I had to get OCIO into ComfyUI for a shot (log for 2D distortions without losing pixel detail), and it wasn't there. So I used the tool to build a full OCIO node pack, 8 nodes, and shipped it the same day: github.com/SlavaSexton/ComfyUI-OCIO. Custom nodes on demand, the pipeline-TD case exactly.
This is fantastic! My Hermes agent had this in the newsletter it makes for me and was already using it to create the newsletter graphics! Well played, Comfy!
Way better than having some aggregate site constantly wobbling direct control of the pipelines we all rely on for actual work.
I like the creative-tool angle here because it makes MCP feel less abstract. Once an agent can reach a visual workflow, the important boundary is no longer just "can it call the tool", but what kind of creative state it can alter, save, overwrite or publish. That is where tool access becomes authority, not just integration.
That sounds promising. The creative-state angle is the one I am watching most closely: once agents can alter a workflow rather than just suggest one, the evidence and undo story become part of the product.
Guys you are too late 3 years ago i used your endpoint locally to fire the flow using the prompt endpoint exactly if i recall and build and amazing site that generates photos and videos . Not sure what your products team were doing, now a days after higgsfield integration and alot of other tools integrated with claude and other llms you can with mcp, this good but kind of late and i believe no one will rebuild his setup and integrations for comfyui !!! As it's not only mcp its full flows in comfy it self that's need a real experts, my suggestion is to publish your api swagger and pre defined flows that's really work and generates good quality output and link all together and sell it 😀 thanks comfy team
Thanks for the comment Nour! Agreed that this could have come earlier in this fast moving world, but here we are 💪🏼 Cleaned up cloud API docs are coming soon along with a bunch more developer tools; looking forward to seeing what you and others build with all of the things we're working on!
Hi Behram! Currently this MCP is only available for cloud users. We are working on solutions for local installations as well but this one should help superpower your cloud account if you are using it. 🚀
As an educator my resources for buying subscriptions for me and students are negligible. I have a 20gb gpu and my students have 12 gb. To cope with advances in AI i have to learn through local resources only. So im looking forward to your local implementation.
Cheers and thank you for building such an amazing toolset .
Congrats, awesome to see this ship. We clearly landed on the same idea from opposite ends: you took it to the cloud, I built it offline-first. ComfyUI - Agent Kit gives the same full natural-language control, where the agent assembles the entire workflow, generates images, video, and audio, and even builds custom nodes (V3 API) for your exact task, across Claude Code, Codex, Gemini, and Qwen. It ships near-daily and folds in community fixes and workarounds by hand, often before they reach an official build, so the agent routes around broken paths before it builds. The twist: it runs entirely on your own machine, private and NDA-safe, and still reaches online models when you want. Cloud for scale, local for control.
github.com/SlavaSexton/ComfyUI-Agent-Kit
Hi Slava! Thank you for sharing (and contributing to the overall ComfyUI community), I just recently came across your repository and plan to dig into it over the next couple of weeks. Sounds very cool!
Matt,
That automated custom node creation I mentioned is already done, v2.0.1
It came out of a real need: I had to get OCIO into ComfyUI for a shot (log for 2D distortions without losing pixel detail), and it wasn't there. So I used the tool to build a full OCIO node pack, 8 nodes, and shipped it the same day: github.com/SlavaSexton/ComfyUI-OCIO. Custom nodes on demand, the pipeline-TD case exactly.
Thank you, Matt! 💪🏼
I’m still tuning my tool to automate ComfyUI custom node creation for pipeline TD experts in the VFX industry (except workflows)
I'm from vfx. 27 years and counting. Thanks so much for doing this . Starred the repo.
Cheers
This is fantastic! My Hermes agent had this in the newsletter it makes for me and was already using it to create the newsletter graphics! Well played, Comfy!
Way better than having some aggregate site constantly wobbling direct control of the pipelines we all rely on for actual work.
Thanks Enid! Looking forward to hearing how you continue to use the MCP and what you're building with it.
I like the creative-tool angle here because it makes MCP feel less abstract. Once an agent can reach a visual workflow, the important boundary is no longer just "can it call the tool", but what kind of creative state it can alter, save, overwrite or publish. That is where tool access becomes authority, not just integration.
Stay tuned! We're cooking something up that is coming soon that I think you'll like based on this comment 📈
That sounds promising. The creative-state angle is the one I am watching most closely: once agents can alter a workflow rather than just suggest one, the evidence and undo story become part of the product.
Guys you are too late 3 years ago i used your endpoint locally to fire the flow using the prompt endpoint exactly if i recall and build and amazing site that generates photos and videos . Not sure what your products team were doing, now a days after higgsfield integration and alot of other tools integrated with claude and other llms you can with mcp, this good but kind of late and i believe no one will rebuild his setup and integrations for comfyui !!! As it's not only mcp its full flows in comfy it self that's need a real experts, my suggestion is to publish your api swagger and pre defined flows that's really work and generates good quality output and link all together and sell it 😀 thanks comfy team
Thanks for the comment Nour! Agreed that this could have come earlier in this fast moving world, but here we are 💪🏼 Cleaned up cloud API docs are coming soon along with a bunch more developer tools; looking forward to seeing what you and others build with all of the things we're working on!
some questions
- can i use it on comfy portable or do i compulsory need comfyui desktop
- do i compulsory need a comfyui account . is this only for cloud users
Hi Behram! Currently this MCP is only available for cloud users. We are working on solutions for local installations as well but this one should help superpower your cloud account if you are using it. 🚀
Thanks for the heads up Matt.
As an educator my resources for buying subscriptions for me and students are negligible. I have a 20gb gpu and my students have 12 gb. To cope with advances in AI i have to learn through local resources only. So im looking forward to your local implementation.
Cheers and thank you for building such an amazing toolset .
behram