Skip to content

Releases

Muse Agent v0.1.7

Choose a tag to compare

@CM71103 CM71103 released this 09 Nov 04:31

Full Changelog: v0.1.5...v0.1.7

Muse Agent 0.1.5

Choose a tag to compare

@CM71103 CM71103 released this 04 Nov 15:32

Memory enabled by default. Bundled Python runtime, so OCR and hands work on a fresh machine with no Python install and no pip step. Append-only JSONL session rollouts with two-tier compaction summaries. Supersedes v0.1.4, which failed to start.

v0.1.3.1 — session rollout + fixes

Choose a tag to compare

@CM71103 CM71103 released this 04 Nov 04:04

Append-only JSONL session storage (crash-safe, legacy migration), two-tier session summaries, focus-mode scroll fix. Everything from the Terax redesign: NeuroNoise background shader, SmokeRing emblem, glass panes with opacity slider, bundled Inter Variable. Multi-turn 400 memory fix. 115/115 tests.

v0.1.3, Terax redesign

Choose a tag to compare

@CM71103 CM71103 released this 03 Nov 17:20

Muse Agent 0.1.3, Terax redesign

Local-first personal AI agent. Electron shell + Node agent harness + Python hands sidecar. No cloud relay, no telemetry, no account. Your key, your machine.

The Terax update

A ground-up visual and interaction rework, plus two correctness fixes that were breaking real sessions.

Design

  • Terax design system: grayscale-inversion primary color, lime green kept as the brand accent only (emblem, pips, hero). Full radius, motion, and surface token scale, oklch syntax palette. Light and dark themes rebuilt on it.
  • Ambient WebGL shaders: faithful ports of Paper Designs shaders. NeuroNoise as the animated background behind frosted-glass panes, and a SmokeRing smoke effect on the hero emblem. Zero dependencies, no CSP relaxation, respects prefers-reduced-motion.
  • Frosted-glass panes with a live opacity slider in Settings, Appearance. Drag to tune how much of the shader reads through the glass.
  • Typography: Inter Variable + JetBrains Mono, bundled as woff2. No more system-serif fallbacks.
  • Resizable sidebar: drag the handle (180 to 480px), double-click to reset. Spacious session/project sidebar with clear selection highlights.

Fixes

  • Multi-turn 400 error fixed: project memory was spliced into the conversation as a second system-role message. Providers that require system messages only at the start rejected every turn after the first. Memory now rides inside the leading system message behind a stable delimiter.
  • Sidecar spawn errors now say what actually failed (ENOENT code, exe tried, MUSE_PYTHON override hint) instead of a generic guess. The py-launcher fallback is supervised correctly.

Also in this cycle

  • Image input via OCR (paste/drop, hands sidecar)
  • LLM-summarizing context compaction + /compact
  • Claude Code-style permission modes (ask / accept_edits / plan / auto)
  • Update checks (GitHub Releases, notify-only)
  • Sidecar restart-loop fixes, port-conflict dedupe

106/106 tests passing.

Download

  • Muse Agent Setup 0.1.3.exe: NSIS installer, Windows x64
  • MuseAgent-0.1.3-portable.exe: single file, no install

First run seeds config at %USERPROFILE%\.muse-agent\config.toml.

Note: no code signing (hobby project). SmartScreen will warn. "More info, Run anyway".

Setup

  1. Run the exe
  2. Edit %USERPROFILE%\.muse-agent\config.toml. Presets inside for OpenAI / DeepSeek / OpenRouter / Ollama / LM Studio
  3. Set your key: set MUSE_LLM_API_KEY=sk-...
  4. Chat. Every shell call asks first. Approvals are the security model. Hands (screenshot/mouse/keyboard) is off by default. When on, every action still asks.

Build from source: npm install && npm run dist (Node 18+, Electron 45).

MIT license. It can do anything you approve. Approve carefully.

MuseAgent-0.1.1-lime-portable.exe

Choose a tag to compare

@CM71103 CM71103 released this 25 Oct 13:41

Muse Agent 0.1.1, the lime era

First public release. Local-first personal AI agent. Electron shell + Node agent harness + Python hands sidecar. No cloud relay, no telemetry, no account. Your key, your machine.

What was in it

  • Subagents: the agent can delegate self-contained subtasks to fresh child loops (spawn_agent). Own context window, turn budget, inherits project folder, deny rules, approval gates. Children can not spawn children.
  • MCP servers: plug in Model Context Protocol tool servers (stdio) via config.toml. Tools surface as mcp__<server>__<tool>, every call gated by approval.
  • Host bridge: loopback-only HTTP API on the live agent (/send, /status, /cancel, /sessions). Drive Muse from scripts, hotkeys, other tools. Bearer-token auth, 127.0.0.1 only.
  • Tray + notifications: background presence with idle/working tray state, cancel/new-session/quit. Desktop notifications on approvals and turn completion when unfocused. Hands sidecar supervised, auto-restart on unexpected exit.
  • The lime look: lime theme, spark emblem across app and installer icon, particle-reconstruction boot splash, hover replay on the mark.
  • Hardening: Electron IPC senderFrame validation, origin-bound preload keys, navigation hardening, sentinel policy with scope grants, question tool, cancel support.

89/89 tests passing at the time.

Download

MuseAgent-0.1.1-lime-portable.exe, single file, no install, Windows x64. First run seeds config at %USERPROFILE%\.muse-agent\config.toml.

Note: no code signing (hobby project). SmartScreen will warn. "More info, Run anyway".

Setup

  1. Run the exe
  2. Edit %USERPROFILE%\.muse-agent\config.toml. Presets inside for OpenAI / DeepSeek / OpenRouter / Ollama / LM Studio
  3. Set your key: set MUSE_LLM_API_KEY=sk-...
  4. Chat. Every shell call asks first. Approvals are the security model. Hands (screenshot/mouse/keyboard) is off by default. When on, every action still asks.

Build from source: npm install && npm run dist (Node 18+, Electron 45).

MIT license. It can do anything you approve. Approve carefully.