Changelog

Agents get schedules and folder watches

Agents now run on cron-style schedules or fire when a watched folder changes, with every run written to a plain-text transcript in ~/ax/runs. Tool grants moved into the sidebar so an active shell or file permission is always visible while it is live.

Engine

Vulkan backend for AMD and Intel GPUs on Linux and Windows. Model swap while a conversation is open no longer re-tokenizes history, which makes trying two models on the same question instant.

The local API server

One toggle exposes loaded models at localhost:4891/v1 with OpenAI-compatible chat, completion, embedding, and model-list routes, with an optional bearer token. Editors and agents that already speak the dialect connect with a base-URL change.

Vision and speech land

Llama 3.2 Vision and Whisper large-v3 join the library: drop a screenshot into chat, or transcribe audio fully offline.

The model library

Curated open-weight models with per-quantization memory footprints, publisher license notices before first download, and one shared weights folder across conversations. GGUF import and Hugging Face URL paste for everything else.

Reasoning models

DeepSeek-R1 distills with visible thinking traces, collapsible in the transcript.

AX exists

First internal build: local chat on Apple Silicon and NVIDIA hardware, hardware profiling that recommends a model which actually fits, and history stored as plain local files. No account, no telemetry, network cable optional. The three rules were set here and have not moved since.