A major pass focused on letting Axiom run against your own hardware, plus hardening the systems that keep long conversations and multi-agent runs coherent.
Bring your own inference server — point Axiom at any self-hosted OpenAI-compatible endpoint (e.g. your own Ollama box) as a first-class model option, in both Normal Chat and the Workplace Council. A new Hybrid Local mode blends local reasoning with your self-hosted backend, with tool-calling and context-window budgeting that correctly scale down for smaller self-hosted models instead of assuming a large cloud context window.
Context compaction, fixed for real — the automatic long-conversation summarization system (Normal Chat and Workplace Council) had a bug where summaries could silently vanish instead of actually being used going forward. Both are now correct, with safer fallback behavior and clearer logging when something needs attention.
MCP connectors & cloud agentic tools — Gmail/Drive, GitHub, and Todoist OAuth connectors, a new Settings Connectors UI, and Workplace GitHub integration.
Security — the OpenRouter API key is now encrypted at rest via Windows DPAPI.
Reliability & polish — exact (not estimated) OpenRouter token usage, diff-aware artifact Critic reviews, a first-run default model downloader, native markdown rendering in Workplace cards, background council run completion notifications, chat search/rename/pinning, bounded rendering for long chats, Builder vision support in Workplace, a Settings backend log viewer, transient chat failures now surface in a banner instead of failing silently, and a new pure-logic unit test suite.
Runs local-first on Windows. Unzip and run, or update in-app.