Eaon 2026.3.2

Added

• Free Week — 7 days of every hosted model, free, one click, no
  account, no card. Bound to this Mac (hardware fingerprint), so
  reinstalling the app recovers your existing trial instead of
  starting a new one. Start it from onboarding, Settings → Providers,
  or the new one-time popup nudge for installs that already dismissed
  onboarding before this existed.
• Settings → General → Gifts — a stable place to see and redeem what's
  currently available, showing the live claimed/remaining count and
  redemption deadline for the Free Week.
• Eaon CLI now tells you when the copy installed by a previous version
  of the app is out of date — a badge and Update button in the CLI
  hub, plus a status line in General. Updating re-copies the newer
  bundled build without touching your CLI config or sessions.

Changed

• The bundled Eaon CLI now checks Eaon Desktop's own copy on launch
  and prints a one-line heads-up in the terminal when it's behind.

Eaon 2026.3.1

macOS-only update over 2026.3.0.

Added

• One-click Eaon CLI install (Settings → General → Eaon CLI) — a
  ready-to-run copy now ships inside the Mac app; Install copies it
  to ~/.eaon/cli-app and links a global eaon command, no npm or
  network needed.

Eaon v2026.3.0

Cross-platform builds of Eaon.

• Windows: download the .exe installer. (Beta) (Need help building the Windows version of the app) If you want access, contact support@eaon.dev or sanscreates@eaon.dev or sanscreates for Discord
• Linux: download the .AppImage (portable, runs anywhere) or the .deb (Debian/Ubuntu).

Heads up — these builds are not code-signed yet. On Windows,
SmartScreen shows a blue "Windows protected your PC" box: click
More info → Run anyway. This is expected for unsigned apps
and goes away once the app is signed.

What's Changed

• Improve README quick start by @morriszdweck in #1

New Contributors

• @morriszdweck made their first contribution in #1

Full Changelog: v2026.2.0...v2026.3.0

Eaon 2026.2.0

Added

• MCP & Skill integration — connect any custom MCP server with a pasted
  token, and install a Skill Library from GitHub or your own local Claude
  Code skills. Invoke any enabled skill directly in chat with
  /skill-name.
• A rebuilt memory system — Eaon can learn from your conversations over
  time (with consent), remembering specific things you've shared instead
  of only static facts, and can learn from a file you pick or from a
  connected plugin's results.
• Live code streaming with syntax highlighting — watch code appear in
  color as the model writes it, and the chat no longer yanks you back to
  the bottom while you're scrolled up reading.
• A per-model CPU/GPU control for downloaded Hugging Face (llama.cpp)
  models — force CPU-only, force max GPU, or leave it on auto.

Fixed

• Chat mode not actually writing code to disk when asked.
• The sidebar re-sorting your chats out of order just from clicking
  between them.
• Several Hugging Face model download/run issues: a hard crash on
  Gemma-family models, misleading "try a different quantization" advice
  when a model's whole architecture isn't supported, and a truncated
  error log that hid the real reason a model failed to load.
• Deleting a downloaded model reporting success without actually freeing
  disk space.
• A confusing "wait for models to load from the Aqua API" message
  appearing even when the real issue had nothing to do with Aqua.
• pip's "externally-managed-environment" error when Eaon's agent tries
  to install a Python package.

Eaon 2026.1.8

Added

• Image generation — use an image model over the API (Aqua's hosted
  models, or a BYOK cloud key) or a local one running on this Mac
  (Automatic1111-compatible servers like DrawThings/ComfyUI, or an
  Ollama-served diffusion model), with zero extra setup. Settings →
  Image Providers.
• A model attribution header (name + logo) on every reply, so it's always
  clear which model actually answered.
• Multi-step agent replies (a tool call followed by more text) now render
  as one continuous message instead of restarting the header and typing
  indicator for every step, with a live "what's happening right now"
  indicator between steps — including a subtle per-letter wave animation
  on the "Thinking…" text.
• A "Thinking" dropdown under any reply that used real reasoning
  (DeepSeek-R1, QwQ, and other reasoning models served locally through
  Ollama, or DeepSeek's own API) — click to see the model's actual
  chain-of-thought, collapsed by default instead of dumped as raw
  <think> tags into the reply.
• "Always allow tool calls" (Settings → Privacy, on by default) — code
  execution and connected-plugin (MCP) tool calls run without asking
  each time. Desktop Control still confirms every action regardless,
  since it can move the mouse and type on your behalf.
• Local API Server (Settings → Local API Server, off by default) — run a
  local, OpenAI-compatible server on this Mac that any external tool
  (a script, a coding CLI, another app) can point at, routed through
  whichever backend — Aqua, a BYOK key, or a local model — actually
  serves the requested model. Streaming and non-streaming both work.
  Bound to this Mac only; requires an API key by default.

Fixed

• A stray rounded line could appear along the top of the sidebar, from a
  system titlebar decoration view rendering at the wrong size once the
  window's traffic-light controls were repositioned.
• A downloaded local image-generation model (e.g. from Ollama) failing
  with "does not support chat" now works — image models need a different
  endpoint than chat models, handled automatically.