Skip to content

v3.0.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 27 May 04:04
v3.0.0
f2ae02d

Claude Code Studio v3.0.0

The multi-model release. Three months of work landing in one push: a
33-model curated catalog of local + API models with hardware-aware
recommendations, a file directory navigator, accurate per-bucket
resource monitoring, a --dangerously-skip-permissions toggle, an
in-app Danger Zone for Reset / Uninstall, and a long list of red-team
fixes from beta testing.

Cross-platform: Windows (NSIS), macOS (DMG), and Linux (AppImage,
.deb, .rpm) all built from the same source.

Download

OS File Notes
Windows Claude-Code-Studio-3.0.0-Windows.exe One-click NSIS installer. Bundles Node + Claude CLI.
macOS (Apple Silicon) Claude-Code-Studio-3.0.0-Mac.dmg Drag to Applications. First launch bootstraps Node + Claude CLI.
Linux Universal Claude-Code-Studio-3.0.0-Linux-Universal.AppImage Portable single-file build. Works on any distro.
Linux Debian / Ubuntu Claude-Code-Studio-3.0.0-Linux-Debian.deb sudo apt install ./Claude-Code-Studio-3.0.0-Linux-Debian.deb
Linux Fedora / RHEL Claude-Code-Studio-3.0.0-Linux-Fedora.rpm sudo dnf install ./Claude-Code-Studio-3.0.0-Linux-Fedora.rpm

What's new in v3.0.0

Multi-model catalog

  • 33 curated local + API models (Qwen, DeepSeek, Llama, Gemma, Granite, Phi, Mistral, embeddings) — see docs/MULTI_MODEL.md for the full list and selection criteria.
  • Hardware-aware recommendations — auto-detects your tier (toaster / low / mid / high / workstation) based on RAM + VRAM and surfaces the right defaults.
  • Project-aware suggestions — scans your cwd for package.json, pyproject.toml, Cargo.toml, etc. and weights recommendations by detected role (frontend / backend / data / systems / mobile / devops).
  • In-panel terminal viewer — see the launched model's PTY output without leaving the Models tab.
  • Pop-out windows — separate BrowserWindow per model for side-by-side workflows.
  • First-run picker — auto-opens after install with top 3 recommendations for your hardware; pre-pulls them via Ollama.
  • Disk-quota check — warns before pulling if free disk space is less than 1.5× the model size.
  • License disclosure — restricted-license models (Llama, Gemma, BigCode) prompt before launch.
  • Add custom model — form to register your own model with the catalog.

File directory navigator (new sidebar panel)

  • Lazy-loaded tree, one folder level per IPC call. Opening a project with node_modules doesn't freeze the renderer.
  • Recent projects list, persists across launches.
  • Show / hide dotfiles toggle.

Resource monitoring rebuilt

  • Three separate buckets: Claude PTYs, model PTYs, Ollama daemon. Pre-v3 the "Claude" gauge silently aggregated everything spawned via PtyRegistry — accurate numbers only when you weren't running local models.
  • O(n) process-tree walk (was O(n²)) — handles big multi-model setups smoothly.
  • Ollama daemon detection via process-name scan — picks up the persistent daemon + its model-loader children even when no model PTYs are spawned by Studio.

Cost accuracy

  • Rates updated to current Anthropic pricing (Haiku $1/$5, Sonnet $3/$15, Opus $15/$75 per 1M tokens). Verified against anthropic.com/pricing.
  • Local models are free + not counted — disclaimer made explicit.

GitHub error handling

  • 401 / 403 / 404 / network errors get one-line actionable messages instead of raw stack traces. Rate-limit errors include the reset time when present.

Settings additions

  • --dangerously-skip-permissions toggle in Claude CLI section. When on, the Claude CLI launches with permission prompts bypassed automatically. Applies only to Claude PTYs, never to local models.
  • Danger Zone with two buttons:
    • Reset user data — wipes 18 known JSON / JSONL files (settings, history, model registry, debug logs); leaves Chromium profile state intact.
    • Uninstall Claude Code Studio — Windows spawns the NSIS uninstaller; macOS opens Finder at /Applications with drag-to-Trash instructions; Linux detects install format and surfaces the right apt / dnf / rm command.
  • Status bar shows current git branch + dirty indicator. Polls every 30s + on window focus.
  • Single source of truth for version — title bar, status bar, and About row all read from app.getVersion(). Pre-v3 each was hardcoded and drifted (the v2 build shipped with v1.0.0 in the title bar).

Sign-in flow

  • Sign In button now sends /login (Claude's in-session slash command) instead of claude login (which the running Claude session interpreted as chat text). Caught during beta testing.

NSIS installer (Windows)

  • Ollama is no longer bundled in the installer. Beta.1 silently downloaded + installed Ollama (~2 GB) during install with no progress UI; users reasonably thought it was stuck. v3 detects Ollama at runtime instead and offers a one-click install link via the FirstRunPicker.
  • Uninstaller prompts to also remove the user-data JSON files (defaults to No so a planned reinstall keeps settings).

Auto-updater

  • Skips beta builds entirely so beta testers don't get a 404 stack trace from electron-updater looking for latest.yml on a release that doesn't have one. Stable v3.0.0 builds auto-update normally.

Upgrading

From v1.0.0 — uninstall via Windows Settings → Apps first. v1 used Squirrel; v3 uses NSIS. The two installers don't know about each other.

From v2.0.0 — Windows: uninstall via Settings → Apps (the v3 installer's "remove userData?" prompt will offer to preserve your settings). macOS/Linux: drag to Trash / apt remove and reinstall.

From v3.0.0-beta.x — same uninstall-and-reinstall pattern. Settings are forward-compatible.

Known limitations

  • Per-provider API keys (OpenAI, Gemini, OpenRouter) — only Anthropic auth is wired. Other API models in the catalog need manual claude-style env-var setup.
  • Model comparison view (run same prompt against multiple models) — not built. Workaround: use pop-out windows.
  • Embedding-RAG over past sessions — catalog includes embedding models (Qwen3 Embedding, BGE-M3) but the indexing pipeline isn't built.
  • VRAM per loaded model — Resource panel shows RAM by bucket but not VRAM per model (requires vendor GPU SDKs).
  • macOS code signing — unsigned build. First launch may show "App is damaged" — right-click → Open → Open. To fix permanently, set up Apple Developer Program ($99/yr) + osxSign config.

Credits

Multi-model catalog research compiled May 2026 against:

  • Ollama Library, Hugging Face GGUF index
  • LMSYS Chatbot Arena, HumanEval / MBPP, MMLU-Pro, EvalPlus, BigCodeBench leaderboards
  • Independent reviews: InsiderLLM 2026, RTX 5070 benchmarks (Pooya Golchian), index.dev reviews

Built with Electron 42, React 19, Vite, node-pty, xterm.js, Octokit.

Assets 7
Loading