An AI of
your own.

Under the hood, Proxima is built on vega, an Erlang-style supervised runtime for agents.

The whole operation, in one binary.

Every agent gets an inbox, a budget, tools, memory, and a supervisor that restarts it when it fails — with the dashboard watching it all. Already running Ollama, LM Studio, llama.cpp, or vLLM? Proxima finds it and uses it; a box on your tailnet counts as local. Never heard of those? You never need to — the first run (§5) sets everything up.

The managed runtime is a Go port of hermes-agent's local_runtime (MIT, Nous Research). The measured constants are theirs; the supervision is Proxima's.

Budget math, not vibes.

Before anything loads, Proxima reads the GGUF header, probes your VRAM and RAM, and does the arithmetic. A model that won't fit is refused — a physics refusal, stated plainly — instead of thrashing your machine. Models that do fit get a context ladder (64K → 96K → 144K), per-model launch policies, a supervised llama-server with crash restarts, and idle VRAM reclamation when you walk away.

Any GGUF on Hugging Face.

The curated catalog is the paved road, not a wall. Point at any of the thousands of GGUF repos and Proxima streams just the header — a few megabytes, never the weights — runs the same physics check, then downloads. The first managed boot runs a tool-call smoke eval, and the verdict is stamped per model: you know whether this model, at this quant, on this hardware, can actually drive an agent.

Pairing is gated by a per-machine token, and the phone talks straight to your machine over your own network. The desktop dashboard stays on loopback, untouched.

In your pocket. Still on your machine.

Scan one QR with the Proxima app and the phone takes over: chat with your agents, watch a first-run download land, and answer tool approvals from the couch. An agent doesn't run a command on your machine until a human says yes — terminal or phone, whichever answers first.

Headless runs never auto-download. The offer only appears on a controlling terminal, and nothing moves without a yes.

A fresh machine is one keystroke away.

No model server, nothing staged? Proxima offers to set itself up: one y downloads a small, chatty starter model (Qwen3.5 4B, ~3.6 GB) plus the llama.cpp runtime, and drops you straight into chat with your agents.