Project · AI · local-first
Phrouros
A team of personal AI agents that live on a USB stick. Plug it into any laptop, run one program, and your agents wake up — each with its own personality, its own memory, and its own area of expertise. Everything you say and everything they remember lives on the USB. Pull the stick out and nothing of yours stays on the host machine.
The name Phrouros (Greek: the sentry, the guard, pronounced /FROO-ross/) reflects what each agent is — a watcher placed on your hardware to guard your data and your conversations. Sentinels, not employees.
The promise
Five things that aren't true of most AI products:
- Your data never leaves the stick unless you say so. No upload, no telemetry, no "we'd like to send a copy to our servers."
- No installation on the host computer. The agents run from the USB drive and clean up after themselves.
- The agents that handle your private stuff can't reach the internet. The system blocks the tool calls before they happen, not after.
- It runs on a modest gaming laptop. Designed around a low spec (~8 GB graphics card). Big cloud models are an optional extra you control.
- Every agent is a real character. Pick someone you admire — fictional, historical, alive — and the system builds a personality from what makes them themselves.
Who's on the team
You start with seven agents. Each has its own personality and its own job. You can rename them, give them different characters to play, or create new ones.
| Agent | Role | What they do | Internet? |
|---|---|---|---|
| Atlas | The conductor | Routes what you want to the right specialist; follows up | Yes, gated |
| Mnemo | Archivist | Manages your encrypted notes and journals; the only one that reads your vault | Never |
| Doc | Health & journal | Symptoms, sleep, mood, doctor-visit prep | Never |
| Gen | The polymath | Everyday questions, learning, casual conversation | Yes |
| Hex | The engineer | Code review, refactoring, debugging | Yes (no vault) |
| Tal | The accountant | Picks which model to use per task; tracks spend | Yes |
| Mux | The companion | Stories, games, creative play, banter | Yes |
The two marked Never — Mnemo and Doc — are physically blocked from any tool that could send data out. That's enforced by the system, not by asking them nicely.
What works today
Proven, tested, runnable — each with a captured proof log:
- Visual-novel chat window rendering streaming responses in real time
- Seven pickable agents; click into a one-on-one chat
- Thermal display with color-coded warnings (cool / warm / hot / critical)
- Live view of which models sit in your graphics card; one-click unload
- Cloud agents physically blocked from the vault — verified end-to-end (a canary password in a test vault never leaked)
- Smart routing: code questions → code model, "remember this" → journal model, normal chat → main model
- First-run setup that probes hardware and asks worldview questions
- Memory that recalls across sessions in under 100 ms
- Agents handing work to each other; create your own agent at runtime
- Group chat that picks who responds based on what you said
On the roadmap
Designed with clear paths, deferred only to keep the first release shippable:
- Decentralized data sharing — a journal entry written on one USB shows up on another the next time it's online; private agent/asset sharing between users, no servers.
- Emotion-aware agents — tracks how each agent's state shifts over conversations, and how you feel about subjects over time.
- Real-time voice — agents speaking in voices you choose, with an explicit consent flow for any cloning.
- Generated portraits so the contact list looks distinctive instead of initials in a box.
- Prompt-injection shield — a small model that screens untrusted incoming text before it reaches the agents.
Hardware floor
| Tier | Hardware | What you get |
|---|---|---|
| Minimum | 8 GB RAM, integrated graphics | Bootstrap + smallest local models + cloud for heavy work |
| Recommended | 16 GB RAM, RTX 2070 / 4060-class GPU | Mid-tier local models, all agents responsive |
| Comfortable | 32 GB RAM, RTX 4080+ with 12 GB+ VRAM | Multiple models at once, fluid group chat |
USB drive: 16 GB minimum, 32 GB recommended.
What this is not
- Not a replacement for the big hosted chatbots — a 7B local model is not a frontier model. Phrouros uses cloud models through the accountant when it makes sense.
- Not anonymizing you from cloud providers when you choose to use them — it just makes sure your vault never gets sent.
- Not a password manager. The encrypted vault is for notes and knowledge, not credentials.
- Not a closed garden — every personality file, memory, and config is plain Markdown / JSON / SQLite. Your data comes out any time.
Honest current state: a serious engineering project, not a polished product yet. Many subsystems work end-to-end with captured proof; the final packaging pass is what's underway now.