Local-first · one owner · honest about its limits

A tiny software studio that lives on your computer.

FM Studios is a personal virtual startup. You are the manager; a small team of AI employees will pick up development tasks while you stay in charge of every approval.

Today you can run the studio itself — tasks, approvals, projects and recovery. The AI doing the work is the next chapter, and this page says clearly which is which.

Concept artThe 3D office is planned, not built. This is what it is aiming for.

The idea

A studio you run, not a chatbot you ask

Most AI tools answer when spoken to. FM Studios is built around the opposite: standing roles, a task board, and a manager who decides what is allowed.

Works today
Local-first
Tasks, projects and settings live in a SQLite file on the computer you use. Your laptop and your desktop each keep their own studio, and nothing syncs between them.
Works today
You stay the manager
How freely work is handed out is a setting. Even on auto-assign, an employee gets no permission to touch files, run commands, commit or push. Silence is never approval.
Works today
Honest about failure
If the studio closes unexpectedly, work that was running is marked interrupted and reported to you. It is never resumed behind your back.

How it works

From a sentence to a reviewed result

The first three steps work now. The last two need the AI executor that comes next, so until then task progress is moved along by buttons that say plainly they are simulations.

Step 1: Write a task
Describe the work and optionally attach a registered project folder.
Works today
Step 2: Choose who does it
Pick an employee. Their autonomy setting decides whether it is assigned at once or proposed to you first.
Works today
Step 3: Approve or decline
Proposals wait in your inbox until you answer. Risky setting changes need approval too.
Works today
Step 4: The employee does the work
An AI employee works in an isolated git worktree using the model you configured.
Planned
Step 5: Review the result
Read what happened, then accept it, send it back or discard it.
Planned

The team

Roles with stable identities

Each employee is a standing role with a name, a look, a model preference and its own autonomy and permission settings. The preferences are metadata for now: no model is called yet.

Manager
You
Sets the rules, approves proposals, reads the results.
Engineer A
Software engineer
Prefers Claude.
Engineer B
Software engineer
Prefers Claude.
Engineer C
Software engineer
Prefers Claude.
Engineer D
Software engineer
Prefers GPT.
Product agent
Product
Shapes tasks and acceptance criteria.
Designer
Design (optional)
Off by default.

Works today the registry, settings and approvals ·Planned the employees actually doing work

The office

A floor that shows what everyone is doing

The dashboard already works out what each employee should be doing from real task data. The voxel office will draw those same states as small animated characters.

Planned Character states the office will show

Working
Heads down on a running task.
Waiting for approval
A proposal is in your inbox.
Needs attention
A task failed or was interrupted.
Heading home
The studio is closing for the day.

Light on purpose

The scene above is a few dozen vector shapes with no image downloads and no scripts. The real office will load only on its own page, respect reduced-motion settings, and fall back to the plain text view on weak graphics hardware.

Under the hood

Boring, open-source building blocks

Everything runs on free, open-source software on your machine. No paid authentication service, no cloud database, no subscription-only framework.

In use today

Next.js (App Router)
React
TypeScript
Tailwind CSS
Fastify
WebSocket
SQLite + Drizzle ORM
Zod
Auth.js · Google sign-in
Vitest
pnpm workspaces

Planned

Claude API
OpenAI API
Ollama (local models)
Git worktrees
Telegram approvals
3D voxel renderer

AI providers need their own API keys (or a local model). A chat subscription is not assumed to include API access.

Development status

Where the project really is

Built milestone by milestone, with review between each one.

Works today

Create, assign, approve, cancel and re-queue tasks, with a full history
Two assignment policies, per-employee overrides, and separate permission switches
Deferred settings that wait for a safe moment, or for your approval
Register local project folders (checked, never modified)
Recovery notices after an unexpected shutdown
Sign-in with one allowed Google account

Not yet

No AI model is called yet; task progress is moved by labelled development simulations
No commands run and no files are changed in your projects
No Telegram, no authenticator-app second factor, no 3D office

M1
Foundation
Dashboard, team registry, task lifecycle with history, approvals, deferred settings, project registry, crash recovery.
Works today
M1.1a
Public site and private sign-in
This page, and Google sign-in for one allowed account protecting the dashboard, the API and the live connection.
Built, in review
M2
Safe execution
A first executor, git worktrees, and permissions that are actually enforced, still without a paid model.
Planned
M3
First real AI provider
Starting with local Ollama, then Claude and OpenAI through API keys you provide.
Planned
M4
Notifications
Approve or decline from Telegram.
Planned
M5
The voxel office
A small isometric 3D office where your team sits, works and waits.
Planned

A tiny software studio that lives on your computer.

A studio you run, not a chatbot you ask

Local-first

You stay the manager

Honest about failure

From a sentence to a reviewed result

Step 1: Write a task

Step 2: Choose who does it

Step 3: Approve or decline

Step 4: The employee does the work