ComfyUI with no GPU: Meta's AI agent rented one for me

Six minutes forty seconds. There was no terminal on my side at any point.

I gave Muse one paragraph. It downloaded the free kit, unzipped it, read the setup notes, created its own Python environment, worked around a proxy, started ComfyUI with the Krea 2 Turbo model on a rented GPU, made one image, shut the container down, and then went and read its own billing report. On the cached rerun of 2026-10-03, the execution stage took about 95 seconds by Muse's own report, and about 3 minutes passed from the moment I sent the prompt to the image landing. Muse's settled usage figure was about $0.024.

I am not going to call that "free". Modal required a payment card on file before any GPU would start — I tested that on a brand-new account and it refused. What my card was charged for that workspace that month was $0.00, because credits covered the usage, but that is a credit being applied, not a promise about your account. Extra usage can cost money.

• Access to Muse, and a plan that allows it. Muse is Meta's personal agent. For availability and plans, go by Meta's own announcement: at the time of writing it names the United States and Canada, and what your own account actually gets is down to your account. My run used Muse's free usage; Muse also has paid plans. I can't promise yours will match.
• A Modal account, with a payment card on file. This is the part people dislike, so it goes first. GPU access on Modal needs a card even to use the monthly credit — I re-tested this on 2026-10-03 with a fresh account and no card, and the run stopped with Modal's own message: "Please add a payment method to use L4 GPU functions." No GPU ran, nothing was downloaded, nothing was billed. That is Modal's rule, not the agent's.
• The free kit. Same kit as episode 1, and it now carries the fixes this run needed: the environment, the proxy, and the permission notes. You can get it further down this page.

Not affiliated with or sponsored by Meta, Modal or Krea. I use my own accounts and pay Modal directly.

• Execution: about 95 seconds. Muse reported that stage itself on the cached rerun of 2026-10-03 (its own stage timings: ComfyUI ready in about 25 s, image about 50 s, container about 75 s). That is the run being described here, not a full session.
• Send to image: about 3 minutes. From the moment I sent the prompt (10:19 PM) to the image landing (10:22 PM).
• Usage reported by Muse: about $0.024 (its settled figure, reported 10:23 PM: L4 about $0.0173, CPU about $0.0056, memory about $0.0012 — the three add up to about $0.0241). This is a usage figure on the Modal account, not an amount charged to a card.
• Charged to my card for that workspace: $0.00 that month. That came from the Modal billing overview page (total spend this month, credits applied) — a different source from the figure above. The two are separate and should never be added together.
• The very first full session: about 13 minutes (2026-09-30, first recorded frame to finished image). Most of that was me approving things, not compute. That was a cold run; the 95 seconds above is a cached rerun.
• An L4 GPU listed at about $0.80 an hour when I checked on 2026-09-30 — the GPU only. CPU, memory and storage are billed separately, and prices change.

Here is the discipline that produced those numbers. Muse did not quote a price straight away — it said billing hadn't settled, went away, checked again later, and came back with the total and a breakdown. If your agent quotes you a cost in the first minute, it is reading a running estimate, and that estimate moves. Read the settled figure, and say which source you are quoting.

And read this one twice: a monthly credit is an allowance, not a free tier. It applies because a card is on file, it resets, it can change, and it runs out. On my account the usage limit was one I had set myself ($30). I don't know what a brand-new account starts with, so open your own billing page and read it before you run anything.

1. Open the billing page and read two settingsModal's usage and billing page shows a usage limit and a separate, harder spend limit. They are two different controls. On my page the usage limit was set to $30 and the page said it could be raised as high as $42.50 — that $42.50 is the ceiling the page offers, not the value your account starts at.
2. Decide both numbers yourselfI set the usage limit to $30 to match the monthly credit, and left the spend limit at the account's own default rather than inventing a value. I haven't tested exactly what each control stops, so read Modal's page and choose yours deliberately — do not assume either starts at zero.
3. Log in once, in your own browserMuse cannot log in for you and should not try. It hands you a link; you open it and approve. My prompt also forbids printing any token and touching payment or account settings. Read the prompt before you send it, not after.

I'm not claiming any setting guarantees a zero bill. Credits covered my test, and extra usage can cost money.

Download the starter kit from this link: [kit link]
Unzip it, then read SKILL.md and SETUP-CHECKLIST.md.
Use Modal to run ComfyUI with the Krea 2 Turbo model on a GPU and make one image:
"a vintage film camera on a wooden desk by a window, soft morning light, real photograph, 85mm lens".
Rules: wait for my approval before any login or network permission; never print a token; do not touch payment or account settings.
When you're done, tell me what it cost after Modal's billing has settled.

Three lines in there are doing deliberate work, and they are worth stealing for any agent you hand a cloud account to:

1. Get Muse access and confirm your plan allows this.Go by Meta's announcement for availability and plans, and check your own account rather than assuming. Muse is the whole reason there is no terminal on your side.
2. Make the Modal account, add a card, set the limits.A card on file is required for GPU access; then open the billing page and set the usage limit yourself. This is the only step with money in it, which is why it comes before the fun part.
3. Download the kit and give Muse the prompt.The kit is further down this page. Paste the prompt above, with the kit link in the first line. Then step back: Muse works through it on its own screen, and you can watch its task list update.
4. Answer the login and permission questions when they appear.You approve the login in your own browser. On the first session I also got Muse's network-permission prompts, one after another, and I kept choosing "allow once" — it asked six times or more. There is a note in the kit about when "always allow" is reasonable (only for the addresses you have actually checked).
5. Check the picture, then the bill.Look at the image with a critical eye — mine had garbled text on the lens, which the agent flagged before I did. Then wait a few minutes and read the settled usage on Modal's page, not the first figure you see.

If your agent runs in a terminal on your own machine instead, the episode 1 route still applies, and the same kit covers it — including the Windows encoding setting, which belongs to that route only, not to Muse.

• Claude Code — episode 1, Windows terminal, 2026-09-30. It set up ComfyUI and Krea 2 on a rented GPU from one paragraph, and hit a Windows encoding crash and a GitHub rate limit on the way.
• Meta Muse — its own cloud computer, 2026-10-03. The run this page is about: environment, proxy and login walls, described above.
• Codex CLI — 2026-10-04, a cached rerun. Already logged in to Modal with the model weights cached, it made an image in about 2 minutes 33 seconds. Its settled usage for that run was about $0.0251 — a usage figure, not a card charge, and not something to compare agents by.

Not tested: Codex's desktop client and its approval prompts, a first-time Modal login from scratch on any of these, a cold start with no cached weights, the path where the account has no card on file (beyond confirming it stops), and every other agent. The kit says "works with an agent that can read files and run commands", because that is the actual requirement — it does not claim every agent will.

• ≈95 s — Muse's own stage timings on the cached rerun (ComfyUI started in about 25 s, image about 50 s, container about 75 s). It is the execution stage, not the whole session.
• ≈3 min — the screen timeline of that same run, from the prompt going out at 10:19 PM to the image arriving at 10:22 PM.
• $0.024 — Muse's settled report at 10:23 PM, with a breakdown by GPU, CPU and memory.
• $0.00 — the total spend shown on Modal's own billing overview for that workspace that month, with credits applied. Read off the page, not reported by the agent.
• ≈13 min — the first full session, 2026-09-30, from the first recorded frame to the finished image; mostly approval waits.
• $0.0251 — Codex CLI's settled usage for its 2026-10-04 cached rerun, read from Modal's billing for that run. Not a card charge.
• $42.50 — the ceiling the Modal usage page says the limit can be raised to, read on 2026-10-03. What a new account starts with: not checked.

Every run is different. Prompts, model, resolution, hardware and how long the container idles all move the total. Treat these dated figures as what happened on those days, not as what you should expect to pay.