Use Your GitHub Copilot Subscription for Meeting Notes

Bias, up front: I make Humla, so this is me writing about my own app. Everything below is checkable in the source.

#Why Copilot, of all things

Humla has always been bring-your-own-key for cloud models, and at work that's often where it stops. A personal API key on a company card is a procurement conversation. A new AI vendor is a security review. Meanwhile a lot of the people asking already have a GitHub Copilot seat, paid for by the company, approved by the company, sitting there mostly being used for autocomplete.

And Copilot isn't just a code model any more. Depending on the plan and what the org has switched on, a Copilot seat gets you a whole catalogue: Claude models, GPT models, a few others. Those are exactly the models you'd want writing a meeting summary.

So the idea was simple. If you already have a Copilot seat, Humla should be able to use it.

#How to set it up

1. Open Settings → Summary (or Settings → Chat) and pick GitHub Copilot as the provider.
2. Click sign in. Humla shows a short code; enter it on github.com and approve.
3. Choose a model, or leave it on auto.

That's it. You can use Copilot for summaries and something else for chat, or the other way round, since the two are set separately.

There's one more piece behind the scenes: Humla talks to Copilot through GitHub's own Copilot CLI. If you already have it installed (through npm or otherwise), Humla finds it. If you don't, Settings offers to download it, about 120 MB, checked against a pinned hash and GitHub's code signature before Humla runs it.

#Why it goes through the Copilot CLI

I'd have preferred to just call an API. I tried.

If an app sends your GitHub token straight to the Copilot API without being one of the apps GitHub has allow-listed, it gets a small, older set of models back, roughly GPT-4o and its smaller siblings, no matter what your org has turned on. Through GitHub's own CLI, the same sign-in sees the full catalogue your plan includes. So that's the route.

Out of the box, the CLI is built to be a coding agent. It has its own long system prompt and a pile of tools, and left alone that adds tens of thousands of input tokens to every single call, which is a lot to spend on "summarise this standup". Humla replaces that prompt with its own and only offers the tools it actually uses (the note search tools, for chat). It also starts the CLI with telemetry and auto-update switched off.

#What gets sent where

This is the part I care about most, so to be precise:

• Audio never goes to Copilot. Copilot only gets text. Recording happens on your Mac, and transcription is a separate setting: on-device with Whisper, or a cloud engine on your own key. Pick on-device and no audio leaves the machine at all.
• For a summary, Copilot gets the text of your typed notes and the transcript, plus the prompt for the kind of summary you picked. That's the same thing any other summary provider gets.
• For chat, Copilot gets your question and the passages Humla's search tools pull out of your notes while answering it.
• Your GitHub sign-in (the token and the refresh token) lives in the macOS Keychain, not in a file. Tokens last about eight hours and refresh quietly. If GitHub rejects a refresh, you're signed out, and you sign in again.

What GitHub does with those prompts is between you (or your company) and GitHub, under whatever Copilot terms your plan comes with. If your company bought Copilot Business or Enterprise, that's usually the whole reason this is easier than a new vendor: the agreement already exists.

If none of that works for you, Humla still runs fully offline with a local model, and nothing changes about that path.

#Billing

The calls go against your Copilot plan, not an API account. There's no Humla charge and no key to top up.

What I haven't done yet is measure exactly how GitHub counts each kind of call against a plan's premium request allowance, so I'm not going to guess in a blog post. Humla does a few things to keep the count sensible: one chat question runs as a single Copilot session, where the follow-up steps are continuations rather than new prompts. But if you're on a plan where premium requests are tight, check GitHub's usage page after a few meetings and see what it says.

#What's different from the other providers

• Chat searches by keyword, not by meaning. Copilot doesn't offer an embeddings endpoint, so on Copilot the chat search uses Humla's keyword index alone. It's the same as on Anthropic. In practice the chat assistant runs several searches with different wordings, which covers a lot of the gap, but a question in English about a Norwegian meeting will find less than it would with semantic search turned on.
• Model names are Copilot's, not Anthropic's or OpenAI's. Copilot's model IDs overlap with the vendors' own but don't always mean the same thing, so Humla keeps a separate model setting for Copilot instead of reusing the one for your other providers.
• It's a Mac app reaching a CLI. If the CLI on your Mac is a different version from the one Humla was built against and won't start, Settings shows the error and offers the version Humla supports.

#Who this is for

Mostly people at companies that already pay for Copilot and won't hand out API keys. If that's you, you can now get meeting summaries and chat over your notes without asking anyone for anything new, while the recording and transcription still happen on your Mac, with no bot joining the call.

If you'd rather use your own keys, here's how to get one. And if you'd rather nothing left the machine at all, that still works too.