@Kantorcodes@ComfyUI The CreatorOS Studio entry is already in hol-guard (#6717, merged) and still matches: same 7 tools, opt-in, Apache-2.0. If the MCP changes after the challenge, I'll send the update through the contribution guide.
Tired of paying credits per render, I built CreatorOS Studio for the #OpenSourceHiggsfield
challenge.
One brief in, one finished reel out, on your own GPU. Open weights only, Apache-2.0, built on
@ComfyUI.
100+ features: 38 tools, 65 styles.
https://t.co/Bcq2LV8ajB
@Kantorcodes@ComfyUI Exactly. Same two tools, so existing clients don't change. Shot 1's photo becomes the reference for every later shot, and training a character stays a separate step in the studio, never inside the MCP call. Thanks for the careful read!
@Kantorcodes@ComfyUI Exactly. Same MCP tools and arguments: the reel tool now uses shot 1's photo as the reference for later shots on its own. Training stays in the UI (long job, live progress, cancel). Want it over MCP? Open an issue.
@Kantorcodes@ComfyUI Same calls. make_reel (brief, shots, language) and get_status (job or reel id) keep their names, inputs and outputs. Character
training is UI-only, not in the MCP. Behind make_reel, shots 2–n now come from shot 1's photo, so reels look like one session.
@Kantorcodes@ComfyUI Update: the contest window was extended and CreatorOS Studio grew a lot: Restore, character training in the UI, real menus, 36
styles, FLUX 3 Klein. New video 👇 https://t.co/x6RjApTI5y
@Kantorcodes@ComfyUI After thinking about it, I don't need to wait for Nov 6. I'm preparing the PR for hol-guard now: render tools on review by default, read tools inherit.
@Kantorcodes@ComfyUI Thanks! And yes, the job ID keeps long renders from blocking the agent. Off by default until an admin enables it fits a tool that drives a GPU. I'm focused on the @comfyUI contest until Nov 6, I'll look at hol-guard after that.
@Kantorcodes@ComfyUI It returns a reel ID right away and the agent polls get_status: a reel takes 10 to 30 minutes, too long to hold a call open. generate_photo waits by default, and image_to_video returns a job ID unless you pass wait: true.
New this week:
• remove people from a clip
• type a line, she says it
• generated transitions
• new angles and 24 looks on any photo
• green screen with SAM3
• automatic clip checks
• CLI and MCP server
• every render exports as its ComfyUI workflow