Sleeping → Waking → Ready
Break free from baby-sitting Colab browser tabs. Comfy Colab Kit lets you wake ComfyUI on demand on Google Colab L4 GPUs from your phone, laptop, or autonomous AI agent. Protected by your personal Cloudflare Access zero-trust gate with automated idle shutdown.
L4 · ~1.54 CU/h · 28.4 s per 1024² image
Deploy the lightweight gateway on your VPS behind Cloudflare Access, then wake Colab from your browser or AI agent in 3 minutes.
Clone the repository and copy .env.example to set your Cloudflare Access audience and Google Colab session name.
Launch via Docker Compose on your VPS or server. Caddy handles automatic HTTPS and WebSocket streaming.
Visit your custom domain. One click wakes the Colab L4 GPU, syncs the model, and proxies ComfyUI.
# 1. Clone the repository
git clone https://github.com/FireTable/comfy-colab-kit.git
cd comfy-colab-kit
# 2. Configure your environment and secrets
cp .env.example .env
# 3. Start the gateway daemon via Docker Compose
docker compose -f deploy/docker-compose.yml up -d --build
# 4. Open your domain (e.g. https://comfy.yourdomain.com) and click Wake!A lean, zero-trust bridge connecting your personal devices and AI agents with on-demand Colab hardware.
Compare the daily friction of raw Colab notebooks with our zero-trust, wake-on-demand architecture.
Must sit in front of a computer, manually run notebook cells, and baby-sit the browser tab.
Laptop sleep or network blips kill the session. URLs change unpredictably on every restart.
Either unauthenticated public tunnels or sharing your private Google account credentials.
Headless AI coding agents cannot trigger or orchestrate workflows without human manual setup.
Forgetting to terminate notebooks burns through dozens of costly Compute Units (CU).
Wake from your phone, tablet, or laptop with one tap. No open Colab tabs or PC required.
Stable custom URL fronted by Caddy reverse proxy; background gateway persists independently.
Google/GitHub SSO & Email OTP. Mode B defaults to admin-only wake/stop to protect GPU compute, with seamless team sharing when ready.
Claude, OpenClaw, or LangGraph agents can wake, generate, and spin down via REST/MCP.
Monitors idle activity, max session duration, and CU hard caps to automatically release the GPU.
Engineered for friction-free generation with enterprise-grade session governance.
Wake from your phone, laptop, or API. The gate provisions Colab, configures ComfyUI, and proxies it on the same URL.
Multi-IdP login (Google, GitHub, OTP). Default Mode B restricts wake and stop to admins, while Mode A supports collaborative initiator shutdown.
Wake is strictly idempotent. Everyone sees identical live progress, and a second visitor joins the active VM without duplicating cost.
Sessions automatically go back to sleep on idle timeout, max duration limit, or when the allocated CU budget is reached.
Boots straight into ComfyUI App mode: type a prompt, get an image. Switch back to the full node graph anytime with one click.
Headless AI coding agents can wake the gate, trigger image generation, and release GPUs via REST API and MCP protocol.
MCP server integration based on joenorton/comfyui-mcp-server, mounted at /mcp so AI agents can generate images through your gate.
A profile is a single profile.yaml: public model weights, community workflows, and ComfyUI args. Pick one before waking.
Primary profile: OpenMOSS MOSS-Image on a Colab L4, defaulting to MATLOWAI DeBand int8 workflow with 4-step turbo and official quality graphs.
Stable Diffusion 3.5 on a Colab L4, defaulting to the official SD 3.5 Large fp8 graph with TurboX and Medium turbo fast variants.
Qwen-Image 2.1 official templates on a Colab L4, defaulting to 2.1 Turbo text-to-image with int8 weights that effortlessly fit 24 GB.
Transparent metrics measured on Google Colab L4 GPU infrastructure.
Measured on Colab L4 with the MOSS-Image profile (25 GB weights streamed at up to 330 MB/s). Actual figures will vary depending on network and GPU quota.
Common questions about compute, quotas, workflows, and self-hosting.