Background

The idea build tool to let me interface with AI models from my inbox has been sitting on my to-do list for months. And then the other day I saw this tweet from Jeff Weinstein:

Request for AI agents via email! 🤖📥

Email is a surprisingly pleasant medium for AI agents:

  • turn-based, async
  • no new thing to login to
  • manage N, very different agents at once
  • add links+files
  • web+mobile
  • my shortcut keys work
  • good for short or long inputs+replies

I looked in the replies, hoping there would be a slew of options for my very simple use case of “email OpenAI and Anthropic models.” I wanted to use them and cross that item off my list. But I couldn’t find any.

There were a bunch of tools for “improving email with AI” (like email pre-drafters and inbox categorizers), but none for “improving AI with email” (i.e. using email as the interface to existing AI).

So I treated this as a kick in the pants to go hack it together. Here’s what came out of it:

What’s Mentat Mail?

Mentat Mail lets you send emails to your preferred AI model(s) and receive their responses in your inbox. For most use cases, I prefer it to using the default ChatGPT or Claude interfaces.

Especially with the rise of test-time compute / chain-of-thought models, I don’t want to sit around on those sites / apps waiting for a response. Email is a much nicer interface, and I’m in there all day already.

Jeff’s post above outlines the case well. Plus, I can search across all my agent interactions regardless of LLM provider, I can loop other people into the thread easily, and I can even have different agents interact with each other.

Right now, you’ll need to deploy Mentat Mail yourself — it’s pretty simple (and free, other than the LLM API costs).

Here’s the GitHub with instructions. But if you’re potentially interested in a hosted version where you can just sign up for an account and start emailing with LLMs, drop your email on mentatmail.com.

The hosted version of Mentat Mail would be a simple service (sign up, pick your preferred AI model, get an email address to send to, and start emailing with it). But I’ll only do it if there’s enough interest.

Usage examples

Here are a few examples of where I use it…

Chatting with long-running models

Here I am making a request to OpenAI’s o1 model, which can take awhile. I can drop an email, walk away, and when I come back the response is in my inbox:

It also processes multiple requests concurrently, so you can make an o1 request, then fire off some faster requests (e.g. GPT-4o mini or Claude) while the first one is running running.

Image processing

I can attach an image to the email and the AI will parse it:

Referencing the thread

I can loop the AI on an existing thread and have it use the prior context to answer a question:

(Or I can forward the thread to an agent and it’ll reply just to me.)

Agent-on-agent

I can set up two agents to play off each other, both responding on the thread:

What’s next

Well, I may or may not productionize this. Again, drop your email on mentatmail.com if you’re interested, and/or email me with any thoughts.

I’m pretty happy with the service as-is, personally. But if I spend more time adding features, I think those would include:

  • agent should be able to visit links in the email and use them as context
  • broader attachment handling
  • better handling of email formatting (rich text, math, quoted text, etc.)
  • better error handling (e.g. if the user emailing isn’t whitelisted, perhaps they should get some sort of response about that)
  • custom system prompts for each individual model
  • handling voice notes / audio input
  • integrations into other systems (e.g. adding items to my to-do list)

If you have other ideas, let me know!

Looking for more to read?

Want to hear about new essays? Subscribe to my roughly-monthly newsletter recapping my recent writing and things I'm enjoying:

And I'd love to hear from you directly: andy@andybromberg.com