Offgrid Holdout

My small outpost beyond the algorithmic frontier.

Building an assistant made for an audience of one

Nobody asked for this, and that's the point

I built a personal assistant for exactly one user. Me. There's no pricing page, no waitlist, no plan to turn it into a startup. It runs on my workstation, talks to models on my own hardware, and knows about my calendar, my notes, my projects and the handful of services I keep running at home.

It's called Eilidh (say "AY-lee") PA. It was conceived by me with Claude Code as a design partner and a pair programmer. I am not going to sit here and say I programmed this. I designed, guided, and steered Claude Code to eventually produce what I wanted, not what it wanted. This post is partly about the project and partly about something I think is getting overlooked. AI coding tools aren't only for shipping products faster. They also make it realistic for one person to build software that fits their own life, even when nobody would ever pay for it.

What Eilidh actually does

Under the hood it's a small Python service (FastAPI, SQLite, a Click CLI) with a browser dashboard. Nothing exotic. What makes it useful is that it pulls together things that normally live in separate places:

There's also an optional path to a cloud model, which I'll come back to, because how narrow I made it says a lot about the whole project.

Working with Claude Code

I didn't hand Claude Code a spec and walk away. It was more like having a patient colleague who's always available and never gets bored of the question "wait, why are we doing it this way?"

A lot of the value came before any code got written. I'd describe what I wanted, and we'd argue it out. Should availability be something the model decides, or something code works out from rules? (Code. More on that below.) Should the assistant be able to edit my notes? (No.) What exactly should a cloud model be allowed to see? (Very little.) Talking through those questions with something that pushes back and lays out trade-offs is very different from staring at an empty editor and just writing a detailed prompt.

Then there's the building itself. Claude Code wrote a good share of the plumbing: routes, database code, the tool registry, systemd unit files, the CLI. The stuff I understand well enough but would have taken weeks of evenings to type out. That's the honest reason this project exists at all. Without help, Eilidh PA would have stayed an idea in PLAN.md.

One habit mattered more than I expected. Keeping the plan and the reality separate. PLAN.md holds the big vision (voice, mobile clients, email, multi-machine automation). The project overview only describes what the code actually does today. It's easy to let an AI-assisted project drift into describing what you hoped to build. Making Claude check claims against the code kept me honest.

Built around my needs, and my comfort level

When you build for yourself, you get to decide what you're comfortable with. Nobody's pushing for engagement metrics or more data collection. So I could make choices a commercial product probably wouldn't.

Most things stay local. The app listens only on 127.0.0.1. Models run on my own workstation through Ollama. Tasks, settings, conversations and the audit log live in a SQLite file I own.

Code decides the things code is good at. Whether I'm free on Thursday afternoon isn't a judgment call. It's arithmetic on calendar events and rules. Same for overdue dates, notification timing and whether a disk is too full. The model helps interpret what I'm asking and drafts replies. It doesn't get to guess at facts that can be computed.

The model only gets the tools I hand it. Every tool sits in a registry marked green (automatic), yellow (needs my approval) or red (blocked). The chat model sees read-only tools for calendar, tasks, notes and files. There's no shell access and no "restart that service for me."

The cloud is a narrow door. There's an optional "Advisor" that can ask an Anthropic model for a second opinion. It's off unless I turn it on, it has no tools, and it never sees my calendar, tasks, notes or chat history. It gets the one message I typed, after a redaction pass. Redaction isn't perfect, so I still read what I'm sending.

None of these are clever ideas. They're just what I wanted. The point is that I got to have them.

What changes when there's no money in it

Most software we use is shaped by somebody's business model. It has to appeal to millions of people, so it fits nobody exactly. It has to keep you coming back, so it nags. It has to justify a subscription, so features get bundled and data gets collected. None of that is evil. It's just what happens when software has to pay for itself.

Take the money out and the questions get simpler. Does this help me? Do I trust it? Is it worth an evening? Eilidh doesn't need a login screen for millions of users, an onboarding flow or a growth strategy. It needs to know that I care more about my backup status than my step count, and that some of my projects go quiet for months and that's fine.

For a long time, the catch was effort. Building something like this alone meant giving up a lot of evenings, so per# Nobody asked for this, and that's the point

I built a personal assistant for exactly one user. Me. There's no pricing page, no waitlist, no plan to turn it into a startup. It runs on my workstation, talks to models on my own hardware, and knows about my calendar, my notes, my projects and the handful of services I keep running at home.

It's called Eilidh (say "AY-lee") PA. It was conceived by me with Claude Code as a design partner and a pair programmer. I am not going to sit here and say I programmed this. I designed, guided, and steered Claude Code to eventually produce what I wanted, not what it wanted. This post is partly about the project and partly about something I think is getting overlooked. AI coding tools aren't only for shipping products faster. They also make it realistic for one person to build software that fits their own life, even when nobody would ever pay for it.

What Eilidh actually does

Under the hood it's a small Python service (FastAPI, SQLite, a Click CLI) with a browser dashboard. Nothing exotic. What makes it useful is that it pulls together things that normally live in separate places:

There's also an optional path to a cloud model, which I'll come back to, because how narrow I made it says a lot about the whole project.

Working with Claude Code

I didn't hand Claude Code a spec and walk away. It was more like having a patient colleague who's always available and never gets bored of the question "wait, why are we doing it this way?"

A lot of the value came before any code got written. I'd describe what I wanted, and we'd argue it out. Should availability be something the model decides, or something code works out from rules? (Code. More on that below.) Should the assistant be able to edit my notes? (No.) What exactly should a cloud model be allowed to see? (Very little.) Talking through those questions with something that pushes back and lays out trade-offs is very different from staring at an empty editor and just writing a detailed prompt.

Then there's the building itself. Claude Code wrote a good share of the plumbing: routes, database code, the tool registry, systemd unit files, the CLI. The stuff I understand well enough but would have taken weeks of evenings to type out. That's the honest reason this project exists at all. Without help, Eilidh PA would have stayed an idea in PLAN.md.

One habit mattered more than I expected. Keeping the plan and the reality separate. PLAN.md holds the big vision (voice, mobile clients, email, multi-machine automation). The project overview only describes what the code actually does today. It's easy to let an AI-assisted project drift into describing what you hoped to build. Making Claude check claims against the code kept me honest.

Built around my needs, and my comfort level

When you build for yourself, you get to decide what you're comfortable with. Nobody's pushing for engagement metrics or more data collection. So I could make choices a commercial product probably wouldn't.

Most things stay local. The app listens only on 127.0.0.1. Models run on my own workstation through Ollama. Tasks, settings, conversations and the audit log live in a SQLite file I own.

Code decides the things code is good at. Whether I'm free on Thursday afternoon isn't a judgment call. It's arithmetic on calendar events and rules. Same for overdue dates, notification timing and whether a disk is too full. The model helps interpret what I'm asking and drafts replies. It doesn't get to guess at facts that can be computed.

The model only gets the tools I hand it. Every tool sits in a registry marked green (automatic), yellow (needs my approval) or red (blocked). The chat model sees read-only tools for calendar, tasks, notes and files. There's no shell access and no "restart that service for me."

The cloud is a narrow door. There's an optional "Advisor" that can ask an Anthropic model for a second opinion. It's off unless I turn it on, it has no tools, and it never sees my calendar, tasks, notes or chat history. It gets the one message I typed, after a redaction pass. Redaction isn't perfect, so I still read what I'm sending.

None of these are clever ideas. They're just what I wanted. The point is that I got to have them.

What changes when there's no money in it

Most software we use is shaped by somebody's business model. It has to appeal to millions of people, so it fits nobody exactly. It has to keep you coming back, so it nags. It has to justify a subscription, so features get bundled and data gets collected. None of that is evil. It's just what happens when software has to pay for itself.

Take the money out and the questions get simpler. Does this help me? Do I trust it? Is it worth an evening? Eilidh doesn't need a login screen for millions of users, an onboarding flow or a growth strategy. It needs to know that I care more about my backup status than my step count, and that some of my projects go quiet for months and that's fine.

For a long time, the catch was effort. Building something like this alone meant giving up a lot of evenings, so personal software was mostly for people with lots of spare time or a very specific itch. AI tools change that math. The hours that used to go into boilerplate now go into deciding what I actually want. That's a better use of a person.

I think there's a lot of room here. Not apps that compete with products, but small, odd, personal tools: a garden log that knows your particular soil, a tracker for a family member's medications, a dashboard for a hobby only you and six forum regulars care about. Things that would never survive a pitch meeting and don't need to.

What it isn't (yet)

Eilidh is not an autonomous agent. It doesn't control my desktop, read my email, manage passwords, move money or fix broken services on its own. Voice, mobile apps, RSS and multi-machine automation are all in PLAN.md, and none of them exist yet. Some may never exist, and that's fine too.

It also isn't built for anyone else to use as-is. There's no authentication beyond "you're on this machine," so putting it on a network would be a bad idea without real work first. A family or team would need proper accounts, permissions and separation before sharing it.

And AI help doesn't remove responsibility. Claude Code wrote a lot of code, but I still have to understand it, review it and live with it. Saying "the AI did it" is no help when the morning brief stops arriving.

Build the thing only you want

If you've got a workstation, a few services you care about and a vague list of "it'd be nice if" ideas, try building one of them. Not a product. Just something for you.

Start by deciding what you're comfortable with: what stays local, what a model may touch and what code should decide. Then sit down with an AI coding tool and talk it through before you write anything. You'll probably find, like I did, that the hard part was never the typing. It was knowing what you wanted. And that part was always yours. sonal software was mostly for people with lots of spare time or a very specific itch. AI tools change that math. The hours that used to go into boilerplate now go into deciding what I actually want. That's a better use of a person.

I think there's a lot of room here. Not apps that compete with products, but small, odd, personal tools: a garden log that knows your particular soil, a tracker for a family member's medications, a dashboard for a hobby only you and six forum regulars care about. Things that would never survive a pitch meeting and don't need to.

What it isn't (yet)

Eilidh is not an autonomous agent. It doesn't control my desktop, read my email, manage passwords, move money or fix broken services on its own. Voice, mobile apps, RSS and multi-machine automation are all in PLAN.md, and none of them exist yet. Some may never exist, and that's fine too.

It also isn't built for anyone else to use as-is. There's no authentication beyond "you're on this machine," so putting it on a network would be a bad idea without real work first. A family or team would need proper accounts, permissions and separation before sharing it.

And AI help doesn't remove responsibility. Claude Code wrote a lot of code, but I still have to understand it, review it and live with it. Saying "the AI did it" is no help when the morning brief stops arriving.

Build the thing only you want

If you've got a workstation, a few services you care about and a vague list of "it'd be nice if" ideas, try building one of them. Not a product. Just something for you.

Start by deciding what you're comfortable with: what stays local, what a model may touch and what code should decide. Then sit down with an AI coding tool and talk it through before you write anything. You'll probably find, like I did, that the hard part was never the typing. It was knowing what you wanted. And that part was always yours.

Claude Code, AI-assisted development, personal software, local-first, self-hosted, homelab, Ollama, privacy, side projects

⬅ Previous post
Note – 2026-11-14 12:22