camne 0.2.0

camne now works end to end: question in, local model answers, safety checker reviews it, command out.

• Resident llama-server (llama.cpp b10333, pinned) over a UNIX socket inside a 0700 directory; Windows uses a loopback port instead.
• Zero setup: one consent prompt, the download verified by SHA-256 before it is unpacked, and resumed if the connection drops.
• Safety checker ported wholesale from the whatisit adversarial audit — !! BAHAYA on dangerous commands, and camne still never executes anything.
• Model: nl2sh-1.5b. English, deliberately — the Malay model follows the bake-off.
• Measured on real hardware: 0.16 s warm answer, 1 GB download with progress and retry.

Six static binaries — Linux, macOS and Windows on both amd64 and arm64. No runtime, no dependencies, nothing to install alongside them.

Linux / macOS: curl -fsSL https://raw.githubusercontent.com/officialdad/camne/main/install.sh | sh
Windows: download camne_windows_amd64.exe (or camne_windows_arm64.exe) below.

Every asset's SHA-256 is in checksums.txt. install.sh verifies it before installing and refuses on a mismatch.

On first run camne asks permission to download llama-server and the model (~1 GB, once). Answer anything but y and nothing is downloaded. Everything runs offline after that — nothing you type ever leaves the machine.

🤖 Generated with Claude Code