📁
File-aware local AI
RAG is a major part of the plan. I want Ghost to use my documents, notes, project files, and knowledge base as context instead of forcing me to paste everything manually.
Ghost
Ben Merrill local AI project
Local AI passion project
Ghost is my
private AI
control room.
I’m building Ghost for me. It is a local-first AI workspace running on my own hardware, built to explore what happens when chat, agents, local models, voice, files, remote access, and home control start becoming one personal operating layer.
Mac + Linux app
Python build
AI-assisted
Not for release yet
Summarize my Ghost project notes and tell me what actually matters next.
I’ll pull out the product direction, current build state, open risks, and the next sane move.
Also draft a darker app icon direction.
That’s a heavier creative job. I’ll send it to the desktop and keep your laptop responsive.
LocalMy hardware
AgentsFocused helpers
VoiceTalk to it
Why I’m building it.
I got pulled into local AI because the idea is too interesting to ignore. The problem is that most local AI setups feel like a pile of terminals, model names, config files, scripts, and “almost working” experiments. I wanted something that felt like an actual product. Something private, fast enough, flexible, and shaped around how I actually work.
1
I want controlLocal models, my files, my hardware, my setup. Less blind dependency on cloud tools by default.
2
I want one workspaceChat, agents, models, voice, file context, and remote access should not feel like five separate projects fighting each other.
3
I want to learn by buildingGhost is part app, part lab, part obsession. The fun is seeing how far this can go on real local hardware.
project note
honest status
> goal: build a private AI workspace
✓ Mac/Linux desktop app in progress
✓ Python + AI-assisted development
> local models need real horsepower
! not close to public release
> sharing this as a build journal, not a sales pitch
What Ghost is becoming.
This is the shape I’m chasing. Some of it exists now, some of it is rough, and some of it is planned future work.
🧠
Specialist agents
Focused helpers for product thinking, research, coding, writing, planning, review, and cleaning up messy ideas.
🎛️
Local model control
Installed models, defaults, runability, reply styles, and practical model choices in one place.
🖥️
Desktop horsepower
Bigger models, image work, file indexing, and deep research need real compute. Ghost is being designed around that reality.
📱
iPhone remote
The future companion app should connect back to my server so I can chat, dictate, or voice chat with my own local LLM away from home.
🎙️
Voice and dictation
Talk to it, dictate into it, and use spoken replies when typing is not the right interaction.
🖼️
Creative drafting
Image directions, icon concepts, landing page ideas, UI drafts, and creative exploration all belong in the same workspace.
🔒
Local-first by default
The point is not “never use cloud.” The point is ownership, privacy, and choosing when work leaves my machines.
Current alpha screenshots.
These are real alpha screens from the desktop app. Rough edges are expected. That is kind of the point.
Agent builder
Editing a specialist agent in the current alpha build. Focused helpers are a big part of the direction.
Chat workspace
The main chat surface, built around local models, focused work, and a dark UI that does not feel like a web wrapper.
Attach File
Attach files directly in chat and use them as context. Here I dropped AGENTS.md and had Ghost summarize my agent notes — project scope, working rules, and what is in focus right now.
Model control
Installed local models, reply tuning, and runability signals. Still alpha, but the shape is there.
Settings
Current settings for memory, default models, voice input, and performance behavior.
Voice session
Early voice session testing with open mic, dictation, and spoken replies.
These are not polished launch screenshots. They are receipts from the current build: chat, agents, models, settings, and voice. The goal is to show the direction without pretending this thing is finished.
Planned future
Eventually, I want Ghost to control more than chat.
Longer term, the idea gets bigger: home automation, cameras, sensors, local alerts, routines, and a private command layer for my environment. Not cloud-first. Not some rented brain in the middle. My hardware, my home, my AI layer.
Home automation
Lights, routines, smart plugs, climate, scenes, and everyday controls through a local setup.
Surveillance awareness
Camera and sensor context that can help explain what happened, what changed, and what needs attention.
Local alerts
Useful notifications for motion, doors, devices, unusual activity, or automation failures.
Private command layer
A local brain between AI, devices, files, voice, and my own server.
Current build status.
This is a passion project in active development. It is not a product launch, not a public beta, and not something I’m asking anyone to buy.
🐍
Python app for Mac and Linux
Ghost is being built as a real desktop app, not just a mockup. The current build is focused on Mac and Linux first, with the architecture slowly growing around local models, routing, agents, voice, and files.
🤖
AI-assisted build
I’m using AI heavily to build it, but the direction, taste, product calls, testing, and scope control are mine.
⚙️
It needs horsepower
Local AI is not magic. Bigger models, image generation, RAG, voice, and deep research all need real compute. Trying to get my wife on board with some new hardware for Ghost. 🙂 👻
🚧
Nowhere near release
I’m not close to even thinking about a public release. Right now this is my baby, my lab, and my excuse to learn by building.
🧭
No pitch, no ask
I’m sharing it because I think the direction is cool. That’s it. No buy button, no waitlist, no startup theater.
Ghost, for now.
A dark, local-first AI workspace I’m building on my own hardware. Part product idea, part lab, part long-term bet on private AI.