1.5.0 — Your own silicon

1.5.0 — 2026-08-11 — "Your own silicon"

• Magnitude: find out what your machine can really run — locally, privately.
  Pair the cockpit with a tiny host agent (python3 -m magnitude, pure stdlib)
  that profiles your GPU/RAM, benchmarks a curated catalogue of open models
  with real numbers (gen tok/s, prefill speed, watts, €/Mtok), then downloads
  and serves the one you pick with llama.cpp in a single click. One more click
  connects it to the session router through a local Anthropic-compatible
  endpoint: every new session runs on your own hardware — zero cloud, zero
  cost per token. Opt out per instance: SOKKAN_FEATURE_MAGNITUDE=0.
  Named in homage to Magnitude
  by Tom Greenwald and Anders Lie, whose hardware-profiling onboarding
  inspired this feature (independent from-scratch implementation, no code
  shared, not affiliated).
• Magnitude is multi-machine: the tab is a registry of nodes — pair every
  machine you own (the office workstation, the GPU box, a MacBook), each with
  its own token, hardware profile, benchmarks and served model; pick which one
  powers SOKKAN. Per-node endpoint override for remote nodes (shim_url,
  default SOKKAN_MAGNITUDE_SHIM_URL). Existing single-agent state migrates
  automatically.
• One-line install: pairing a machine is now a single
  curl … /api/magnitude/install.sh?token=… | sh. It lays down a standalone
  Python when the host has none (macOS without Command Line Tools — no sudo,
  nothing outside ~/.sokkan), fetches the agent, and pairs. Validated on
  Linux and Apple Silicon; python3 -m magnitude remains the manual path.