I Gave Meta's Muse a Budget. Here's How It Spent It.

Femi Raphael — Fri, 09 Oct 2026 13:18:36 +0000

Disclosure: I work at AIsa, and AIsa is our product. Everything about Muse below comes from my own use, including the screenshots.

Rio estimated the call would cost about three cents. The quote came back at thirteen. So it stopped and asked me.

That was the moment I decided to write this. Not because Meta's Muse hit No. 1 on the US App Store within 10 days of launch, though it did. Because for the first time, I'd watched a consumer AI app go use a paid API on my behalf, and handle the money more carefully than I expected.

If you build agents, or build things agents will call, the interesting part is how that handoff worked.

What happened when I let it spend

Muse lets you connect outside services. I signed in to AIsa once through OAuth, then asked Rio, the assistant, to find keywords a competitor ranks for that we haven't covered yet. That's a real SEO job, and it runs on paid data.

Three things happened that I didn't expect.

It got a quote before running anything. The content-gap analysis came back quoted at $0.132, with a cap of $0.165. Nothing ran until I'd seen that.

It stopped when the price was higher than it guessed. Rio had estimated about $0.03. When the real quote came in higher, it paused and checked with me before going ahead.

The bill came in under the quote. By Rio's count, every call in the session cost well below what was quoted. One plan quoted at $0.60 ended up costing $0.09.

That's what I think an agent should look like. It can use real professional services, and the spending decisions stay with the person. Notice what made it possible on the service side: an OAuth flow the agent could complete, and a price it could read before calling.

Backing up: what Muse actually is

Muse is Meta's personal AI app. It connects to your calendar, your email, and the tasks on your phone. If something isn't finished when you put the phone down, a cloud machine keeps working on it.

I used it for a few days as my actual assistant: scheduling, email, and the SEO research above. Three things stood out before I ever got to the spending part.

You never manage context. No prompt writing, no fresh chats to keep topics apart. I ran unrelated tasks in one conversation and it kept track. I never repeated background I'd already given it. Most AI products of the last two years quietly assumed you already knew how to use AI. Muse adapts to the person instead. I think kids and grandparents could pick it up without help.

Rio makes it feel like a companion, not a tool. A soft, fluffy character sounds like a small design choice. For someone trying AI for the first time, it's the thing that makes them try.

I couldn't dent the free plan. When I took this screenshot after heavy use, I'd used 2% of my weekly limit.

From answering to finishing

What separates Muse from a chatbot is that it acts across apps, and the work doesn't stop when you put your phone down. Instead of opening five apps to get one thing done, you hand it to an agent.

If you sell online, sit with that for a second. Comparing prices, placing an order, tracking a shipment. More of that will happen through an agent. Your shopper might never see your storefront, just a one-line recommendation from their assistant.

Which is probably why Amazon blocked it

According to this Hacker News thread, Amazon has blocked Muse.

I think the logic is simple. If shoppers get used to buying through someone else's agent, a marketplace stops being a place people visit and becomes a backend that agents query. Traffic, customer data, and ad inventory are all up for grabs at that point. Amazon's caution makes sense to me.

It also raises a question for everyone smaller than Amazon: is your business ready to be understood and used by an AI agent?

Three things I'd do now

Make your products agent-readable. Structured, accurate, complete data is what lets an agent recommend you correctly.
Let agents call your service directly. What I saw inside Muse is that one OAuth sign-in plus clear per-call pricing is enough for an agent to use a paid service with confidence. If your API can be authorized and quoted, agents can buy from it.
Decide whether you need your own agent. Platform rules are still shifting. If your customers reorder often or need help before they buy, an agent of your own keeps that relationship in your hands.

Where AIsa fits

Since I work there, keep your skepticism on. But the experience above is the clearest demo of what AIsa is for: a resource layer that agents can use. One OAuth sign-in connects it. Every call is quoted up front with a spending cap, and you pay per call. And the tools are the professional kind: SEO content-gap analysis, go-to-market research, social lookups, financial data. If you're building an agent, start at aisa.one/api.

The question I'm left with

Have you handed a real task to an AI yet, with money on the line? Not a chat. A task. I'd like to hear how it went, and what it was allowed to spend. Comments are open.

How to Plug AIsa into your Hermes Agent in 2 Minutes (Without Rebuilding Your Setup)

Femi Raphael — Thu, 09 Apr 2026 13:51:03 +0000

Most people make Hermes way harder to run than it needs to be.
They get the runtime working, but then they start changing providers, changing model configs, changing endpoints, and before long the setup itself becomes the problem.

The easier approach is to keep Hermes pointed at one OpenAI-compatible layer and switch models from there.

That is exactly where AIsa fits.

Hermes already supports any OpenAI-compatible API as a custom provider. AIsa is OpenAI-compatible and uses a single base URL at https://api.aisa.one/v1, so the integration path is just the normal Hermes custom endpoint flow.

No special adapters
No hacky workarounds

Here are the two ways to set it up.

Method 1: The CLI Setup (Quickest)

The fastest way to get routing is through the Hermes CLI.

Generate your API key on the AIsa dashboard
Run hermes model in your terminal
Choose Custom endpoint
Enter the AIsa base URL: https://api.aisa.one/v1
Enter your AIsa API key (the one you generated from the marketplace dashboard)
Pick the model you want Hermes to call (e.g., qwen-3.6-plus or claude-opus-4-6)

Hermes’ own quickstart supports this flow, meaning you can easily switch providers or models anytime just by running hermes model again.

Method 2: The Config Setup
If you prefer keeping things in code rather than using the setup wizard, Hermes supports the custom provider path directly in your config file.

Open up ~/.hermes/config.yaml and drop this in:
YAML

model:
  provider: custom
  base_url: https://api.aisa.one/v1
  api_key: YOUR_AISA_API_KEY
  default: YOUR_MODEL_NAME

Why do it this way?

The benefit here is simple. AIsa gives you one base URL and one API key for accessing every major model, while keeping the request format strictly OpenAI-compatible.

You also get provider-agnostic usage tracking and unified billing. This means you aren't rebuilding your Hermes setup or managing five different API dashboards every time you want to test if Llama 3 handles a task better than GPT-4o.

A Quick Reality Check on Fallbacks

If you are doing this for reliability, Hermes recently added support for ordered fallback provider chains through fallback_providers in the config file.

NB: here are currently a few fresh bug reports showing fallback issues on the Hermes API server path.

So the cleanest, most stable recommendation right now is:

Start with AIsa as your Hermes custom provider
Get your primary model working stably first
Add the fallback logic later once the base path is solid

This gives you a much simpler setup, better model switching, and way less provider churn while you’re building.

(P.S. AIsa gives new users free API credits, so you can test this exact routing setup on your agent right now without putting in a card).

DEV Community: Femi Raphael