Local models — on the device
A built-in model store that knows your device — it reads your RAM, chip and GPU
and shows what will actually run, with real download counts and a speed estimate that
self-calibrates on your hardware, plus live search across all of Hugging Face. One tap
runs a model entirely offline through llama.cpp, and it drives the agent out of the
box — the app gives each model the chat template it needs and routes it to the CLI
that fits, so it runs shell commands and reads files with no per-model setup, only the brain
local. Google Gemma, Qwen, Llama, gpt-oss and more, each wearing its own brand —
tool-capable ones wear an agent badge, the rest are honest chat/vision models.
The context window it runs them with follows your device rather than a fixed number, so a
4 GB phone gets a real shelf instead of an empty one — and nothing is offered
that cannot start. Downloads are checked against the SHA‑256 their repository
publishes, gated families work with your own Hugging Face token, and a
.gguf you already own imports as a first-class model.
Talk to it. Out loud, if you like.
Tapping a downloaded model opens a chat with it inside the app — no developer
options, no command line, no wireless debugging first. The chat shows what a local turn
really costs: live CPU, RAM and heat where a cloud chat shows a quota, the window the engine
actually granted, and the measured speed of the last answer. Dictation turns your
voice into text on the phone — 99 languages, no Google speech service, and the
recording is deleted the moment it becomes words. An optional search model makes the
conversations on your phone searchable by meaning, across languages: ask in one and
find the answer written in another, indexed only when you ask.
A model on your phone, for your other apps
Android does not isolate 127.0.0.1, so an inference port on a phone answers
every app on it. Conch’s now needs a key — and another app can ask for
one with a single intent, through a consent screen that names who is asking. You decide,
every grant is listed with one tap to revoke, and what it gets is an OpenAI-compatible
endpoint backed by a model running on your own hardware: no account, no quota, nothing
leaving the device. A loaded model stays loaded behind a notification with one button to
unload it, frees itself when idle, and starts gentler when the phone is already hot.
Your phone is a machine too
A real Linux runs on the phone itself and sits in the machine list like any
server — Conch starts it the moment you open it, because there is nothing for you
to connect to a device already in your hand. And an agent, local or on a server, can
drive the phone at adb-shell level through a small helper Conch installs over
your own SSH link: run commands, read logs, take a screenshot. Nothing leaves the
device except to your own server. Once it is up it needs no network of any kind —
drop Wi-Fi, go to mobile data, go to no data at all: the machine keeps running until the
phone restarts.
Ten agents, one app
Claude Code, Codex CLI, Gemini CLI, Grok, Copilot CLI, Qwen Code, Cursor CLI, opencode,
Crush and Continue — each driven through its own real flags, so resume ids, approval
modes and model pickers are per-agent, not a lowest common denominator.
Sessions that survive
Chats are the CLI's own session files on your disk. Resume days later, from a
different network, on a different phone.
Full-text search
Search across every session on every server by what was said, and jump to the message.
Hardware-key auth
FIDO2 / CTAP2 over USB-HID or NFC. Enrol as many keys per server as you like;
one tap connects, and the same auth carries the whole session.
SAFE / AUTO / YOLO
One switch that maps onto each agent's own sandbox and permission flags, so you
decide how much rope the model gets.
Memory editor
Edit CLAUDE.md / AGENTS.md / GEMINI.md — global and per-project — straight on the server.
Claude extras
Subagent browser and editor, plus your own ~/.claude/commands/*.md
slash commands in autocomplete.
Built for one hand
Markdown rendering, copyable code blocks, mid-turn prompt queueing, a real Stop,
and reconnect-on-network-change that doesn't need a tap.
