Run an agent against your own email for a week

If you're skeptical that local-first AI agents are real or useful, here's a one-week experiment that resolves the question without requiring you to build anything custom.

Total time investment: ~30 minutes of setup + 7 days of running. Total cost: $0 (Avery NXR Free Desktop tier).

Here's the experiment.

The setup

Step 1 — Install: Download Avery NXR from avery.software. Install Ollama if you don't have it (the installer walks you through). Pick a model that fits your hardware (the installer recommends based on your RAM).

Time: 10-15 minutes including model download.

Step 2 — Connect Gmail: Use the Gmail OAuth connector to give Avery NXR read access to your inbox. (Important: read-only initially. We'll talk about send permissions later.)

Time: 2 minutes.

Step 3 — Configure two agents:

→ Email Triager: Reads incoming email every 30 minutes. Classifies into Urgent / Important / FYI / Newsletter / Spam. Posts a summary to a private folder or local file. Takes no action.

→ Email Digest: At 6 PM daily, generates a digest of the day's email — what came in, what you handled, what's still open, what you should follow up on tomorrow.

Time: 15 minutes for both, using template starting points.

The week of running

For the next 7 days, just let the agents run. Don't change anything. Don't expand. Just observe.

Each evening, look at the Email Digest. Notice:

→ How accurate was the agent's classification? → How well did it identify what was important vs. noise? → Did anything fall through that should have been flagged? → Did anything get flagged that wasn't worth attention?

What you'll notice in the first week

Day 1: "Oh, the agent classified that newsletter as 'FYI.' Reasonable."

Day 2: "The agent flagged that vendor email as 'Urgent' that I would have missed in the rush."

Day 3: "The agent classified my mom's email as 'Important' (correct) but also tagged it with the topic 'family.' That's a useful organization scheme I hadn't thought of."

Day 4: "I missed a follow-up that the agent flagged. The digest reminded me. Glad it caught that."

Day 5: "I'm reading the agent's classifications more than I'm reading individual emails. I'm getting through email faster."

Day 6: "I trust the agent's classifications enough that I'm skipping the FYI bucket entirely on busy days."

Day 7: "I should give this agent send permissions for the routine replies."

What the experiment teaches you

By the end of the week, you'll know directly:

→ Whether local AI is real. You'll have seen a local model classify 200+ real emails. You can judge accuracy yourself.

→ Whether agents change behavior. Did you actually use the digest? Did your email habits shift? Did things get missed less often?

→ Whether the platform is usable. You configured agents, monitored them, adjusted them (or didn't). Was that experience friction-free?

→ Whether the privacy model holds. Your email content stayed local. Did you trust that? Did anyone outside your machine see your email content (answer: no)?

These are the questions that marketing pages can't answer for you. The week of running answers all of them.

What you can do at the end of the week

Three paths most users take after this experiment:

Path A: Expand within email. Add a third agent that drafts replies to the routine messages the triager classified as "Important - quick reply needed." You become a faster email responder without doing more work.

Path B: Expand beyond email. Apply the same model to other surfaces. The triager pattern works for Slack mentions, support tickets, GitHub notifications, calendar invites.

Path C: Decide it's not for you. Some people complete the experiment and conclude that the agents added complexity they didn't need. That's a valid outcome. Better to know in week 1 than month 6.

The point of the experiment is to GET to one of those decisions in a week.

Why this specific experiment

We pick email because it has the right properties:

→ High volume — you'll see hundreds of agent decisions in a week, enough to judge quality → High frequency — the agent runs many times a day, exercising the platform → Universal pain — every knowledge worker has email problems → Reversible — read-only agents can't break anything → Personal data — you'll genuinely care about privacy, so the local-first model gets tested against your real instincts

A week of email is enough to feel the value (or not).

Why this beats reading marketing pages

Marketing pages describe what a product CAN do. They show the demo case.

The experiment lets you see what the product DOES in your specific context with your specific data.

The gap between "can do" and "does in my context" is where most product evaluation goes wrong. Running an agent against your real email for a week closes that gap.

What about the privacy concern

If you're hesitant to give an AI agent access to your email — appropriate concern, worth working through:

→ With Avery NXR Desktop tier, the agent processes email on YOUR laptop. Your email content doesn't go to Avery's servers. Avery doesn't have servers in this configuration. → The Ollama model that processes email runs locally. Email content doesn't go to OpenAI, Anthropic, or any cloud LLM provider. → The Gmail OAuth gives Avery NXR read-only access. You can revoke at any time from your Google account settings. → The audit ledger captures everything the agent did. You can review.

If after reading that, you're still uncomfortable — listen to the discomfort and don't do the experiment. It's not for you, and that's fine.

If the concerns resolve with the explanation, the experiment is safe to run.

What I'd tell skeptics

"AI agents will be useful eventually but not yet" is a common position. It's based on extrapolation from current cloud LLM agent platforms (Lindy, Relevance AI, etc.) where cost and complexity make agents feel impractical.

The week-of-email experiment tests a different proposition: that LOCAL AI agents are useful now.

You don't have to take our word for it. The experiment is small, cheap, reversible, and informative.

If skepticism is warranted, the experiment confirms it. If skepticism was holding you back from real productivity gains, the experiment reveals that too.

Either way, you know.

→ avery.software — Free Desktop tier. Run the week-of-email experiment. Make up your own mind.