Standalone local AI

Each device can chat without the other. Pairing is optional for sync / cloud.

v0 try path (5 bullets):

1. Phone alone — install APK → Menu → Model setup → download/import .litertlm → Load → Chat with cascade On this phone.
2. PC alone — run desktop → Host → local models → detect → start llama or litert-lm → use → chat (no cloud key required).
3. Pair (optional) — same Wi‑Fi → Host QR (or ctrl-f5 pair) → phone Scan → CLI exits after ack.
4. Model pick — desktop Providers → model browser → select id for next turn.
5. Stuck? — see firewall / soak notes below and soak-v0.md.

Matrix

Device	Without the other	With the other
Phone	LiteRT chat + local personas/conversations	Cascade / sync / knowledge ingest
PC	Local llama / litert-lm / openai key	Mobile as optional device provider

Phone (Expo)

1. Model setup — download or import a .litertlm file, then Load model.
2. Generator chip On this phone — send uses LiteRT only (no PC/cloud cascade).
3. Default provider order is litert → gateway → zen.
4. Pairing (Host QR / ctrl-f5 pair) is only needed for PC cascade, sync, or cloud keys on the gateway.
5. If chat shows unavailable · …, open Model setup from the banner CTA.

PC (Desktop / gateway)

Backend	Default probe	Env
llama-server	http://127.0.0.1:8080/v1	CTRL_LLAMA_BIN, CTRL_LLAMA_PORT
litert-lm	http://127.0.0.1:9379/v1	CTRL_LITERT_LM_BIN, CTRL_LITERT_LM_PORT

Desktop Host → local models: detect binaries / running /models, Start / Stop, then use to switch the gateway provider. Providers panel includes a model browser (GET /api/v1/models).

Install examples (Windows PowerShell):

# After placing binaries on disk:
setx CTRL_LLAMA_BIN "C:\path\to\llama-server.exe"
setx CTRL_LITERT_LM_BIN "C:\path\to\litert-lm.exe"
# Restart desktop so the gateway child inherits env.

This pass does not download binaries in-app — install CLIs yourself and put them on PATH (or set the env vars).

Pairing

1. Desktop Host QR (CSP allows img-src data:) or ctrl-f5 pair.
2. Use copy LAN URL on Host (prefer 192.168.x.x, not 127.0.0.1).
3. Phone Scan / Connect → health + save → POST /api/v1/pairing/ack.
4. CLI pair polls until ack, then exits 0.

If connect fails: same Wi‑Fi, gateway running, firewall open, use LAN IP (not 127.0.0.1).

Windows firewall (one-liner)

Allow inbound TCP for the gateway port (default 7331), or temporarily:

New-NetFirewallRule -DisplayName "ctrl-a gateway" -Direction Inbound -Protocol TCP -LocalPort 7331 -Action Allow

Backup restore smoke (R3)

Once on a working PC:

ctrl-f5 backup create --label soak
ctrl-f5 backup list
# note an id, then:
ctrl-f5 backup restore <id>

Or use desktop backups panel: create → restore. Confirm a workspace/conversation returns.

Parked until v0 is boring (R4)

Do not start these until soak rows 1–5 pass and you use the build daily:

1. In-app binary downloaders for llama / litert-lm
2. HF / full model catalog browser on PC
3. Real Discord / Slack SDKs (Telegram already real)
4. Device↔host incremental sync beyond backups
5. Rust agentd / monetization experiments