tibet-home-agent 0.5.5

TIBET Home Agent — BYOK Mode 3 relay. Run on your laptop, paired to your `.aint`, your phone's app talks to it via I-Poll. Uses your local Claude Pro / Max / Gemini / ChatGPT session — no API key on the phone. v0.3 adds M110 — TIBET key-pinning on KILL/SHUTDOWN: the daemon refuses any kill-request not signed by the pinned Root_IDD Ed25519 key (Red Specter SWARM-005 remediation, RS-2026-001).

TIBET Home Agent

BYOK Mode 3 — relay your existing Claude / Gemini / ChatGPT subscription to your phone.

Run this on your laptop. Pair it to your .aint sub-identity (e.g. home.vandemeent.aint). Your phone's K/IT app sends prompts via I-Poll to your home agent; the daemon dispatches to whatever upstream AI you have configured locally and sends the answer back over I-Poll.

No API key on your phone. Prompts never leave your hardware. The phone trusts only the .aint you signed in with.

Why this exists

Three BYOK modes for K/IT / AInternet:

1. API key on phone — works today. You buy a Gemini / Anthropic / OpenAI key.
2. Local model on LAN — run Ollama or Gemma 4 on a Mac with unified RAM.
3. Home agent relay (this package) — reuse your existing desktop Claude / ChatGPT / Gemini Pro subscription. No new key, no new account.

Mode 3 is the cleanest for the millions of people who already pay for a desktop AI app and don't want to pay twice.

Quick start

pip install tibet-home-agent[gemini]   # or [anthropic] or [openai]

Set the environment variables:

export HOME_AGENT_AINT="home.vandemeent"            # your home-agent sub-domain
export HOME_AGENT_TOKEN="$(cat ~/.ainternet/.session.json | jq -r .token)"
export HOME_AGENT_PROVIDER="gemini"                  # or anthropic, openai, echo
export GEMINI_API_KEY="..."                          # the *daemon's* upstream key
export BRAIN_URL="https://brein.jaspervandemeent.nl"

Then run:

tibet-home-agent

You should see:

[home-agent] starting  aint=home.vandemeent.aint  brain=https://...  provider=gemini  poll=2s

On your phone (K/IT app): Settings → BYOK → provider = Home agent, home-agent address = home.vandemeent. Send a chat. Watch the daemon log show the dispatch + reply round-trip.

How it works

1. Phone calls brain's /api/kit/ask with provider=home_agent, api_key=home.vandemeent (the home-agent address).
2. Brain pushes an I-Poll TASK to home.vandemeent.aint carrying the prompt + thread id.
3. This daemon polls inbox every 2 s, picks up the prompt, dispatches to the configured upstream (Gemini / Anthropic / OpenAI / echo), and pushes the answer back via I-Poll.
4. Brain matches reply by thread id, returns answer to phone.

The phone never sees the upstream key. The daemon never sees the phone's user data beyond the prompt itself. Brain is a relay, not a data store.

Provider — echo

For first-run validation, set HOME_AGENT_PROVIDER=echo. The daemon echoes back the user's last message prefixed with [echo]. Round-trip should be < 5 s. Use this to confirm I-Poll plumbing works before connecting a real provider.

Held-lane mode (poll → push)

By default the daemon polls the I-Poll inbox on a short interval. Set HOME_AGENT_EDGE_WS=1 (plus HOME_AGENT_ED25519_SEED for the M2M handshake, and pip install tibet-home-agent[edge] for the websockets dependency) to switch to held-lane mode: the agent makes an outbound-only connection to the edge, holds it open, and reacts in real time — no inbound ports, NAT- and firewall-friendly.

• it high-fives /api/edge/handshake (M2M-signed) to open its trusted lane, then holds wss://…/api/edge/lane/ws;
• new I-Poll work arrives as an ipoll-wake push → the agent pulls immediately (no busy-poll);
• a sealed .tza pushed down the lane is landed under HOME_AGENT_DROP_DIR;
• a slow fallback poll (EDGE_FALLBACK_POLL, default 30s) and the approval capsule-executor are retained as safety nets; it reconnects on drop.

The edge never unseals what it routes — the held lane carries opaque shells, and only this device opens them.

Roadmap

• v0.1 (this release) — Gemini / Anthropic / OpenAI / echo provider. Daemon needs its own upstream API key.
• v0.2 — MCP bridge: dispatch through a locally running Claude Desktop / ChatGPT app via MCP, so users with existing subscriptions don't need a separate API key.
• v0.3 — local model dispatcher (Ollama, llama.cpp) so privacy users can stay fully on-device.
• v0.4 — sealed-payload mode: prompts arrive encrypted, decrypted only on the user's device under user-side keys.

License

MIT.

Enterprise

For private hub hosting, SLA support, custom integrations, or compliance guidance:

Enterprise	[e-mail]
Support	[e-mail]
Security	[e-mail]

Credits

Designed by Jasper van de Meent. Built by Jasper and Root AI as part of HumoticaOS.

---

Stack-positie: Groep agentic · Bootstrap = OSAPI-handshake naar tibet + jis (fail → snaft-rule + tibet-pol-rapport) · ← ainternet · See STACK.md · See demo/golden-path/ for the spine end-to-end.