Downloads

Platform	Architecture	File
Windows	x64	ClaudeForge-win-x64.zip
Windows	ARM64	ClaudeForge-win-arm64.zip
Linux	x64	ClaudeForge-linux-x64.tar.gz
Linux	ARM64	ClaudeForge-linux-arm64.tar.gz
macOS (Intel)	x64	ClaudeForge-osx-x64.tar.gz
macOS (Apple Silicon)	ARM64	ClaudeForge-osx-arm64.tar.gz

The version (2026.3.916) is encoded in the release tag rather than in the archive filenames — publish.ps1 produces unversioned artefact names so local dev builds and CI builds share one filename pattern.

Linux: Extract, then run the bundled linux-setup.sh to install a per-user .desktop entry so Wayland compositors render the window icon correctly:

tar -xzf ClaudeForge-linux-x64.tar.gz -C ~/ClaudeForge
cd ~/ClaudeForge
./linux-setup.sh
./ClaudeForge

macOS: The binary is unsigned. Use the bundled allow-app-to-run.sh to strip the Gatekeeper quarantine xattr from every file in the extracted directory:

tar -xzf ClaudeForge-osx-arm64.tar.gz -C ~/ClaudeForge
cd ~/ClaudeForge
./allow-app-to-run.sh
./ClaudeForge

---

What's Changed

Fixed

• Skill and agent descriptions were often wrong, blank, or literally >-. The front-matter parser understood plain and quoted scalars, inline and block lists — and nothing else. Every other shape that occurs in real skill files read back as the wrong value. Three shapes mattered, found by surveying all 236 front-matter files under ~/.claude:

  • Folded and literal block scalars (description: >-, description: |). Nearly every skill Claude Code ships writes its description this way, because the prose contains quotes and colons a plain scalar would have to escape. The header token was read as the value, so the Agents & Skills list, the detail pane and the editor all showed >-.
  • Multi-line flow scalars — a value carried across indented lines with no block indicator at all. Nothing follows the colon in that form, so the value read as empty and the prose fell through as unparsed filler. Anthropic's own math-olympiad plugin skill showed "(no description)".
  • Nested mappings (metadata: with indented keys beneath it) were flattened into phantom top-level fields, and writing one re-rendered it at column 0 — silently lifting it out of its parent. A mapping is now consumed whole and re-emitted verbatim.

  All three chomping indicators (-, +, none) and explicit indentation indicators are handled. A description carrying newlines is flattened for the list row, which is a single ellipsised line; the detail pane and editor keep the real multi-line value.

• Editing one description reformatted the whole file. An edited field now keeps the shape it arrived in — a folded description is written back folded, and a value spread over several lines stays spread over several lines rather than collapsing into one very long plain line.

• Saving a Windows skill file rewrote its line endings. Parsing stripped the \r from each line, so a CRLF file whose block scalar came back untouched was written out with LF inside the block and CRLF everywhere else. Opening a file and saving it with no edits left it with mixed line endings.

• Deep links into a tab only worked on one page. Only Agents & Skills could consume a tab segment; every settings page — Hooks, Permissions, General, all of them — silently dropped it and landed on whichever tab it defaulted to, and never persisted a tab, so place-keeping could not bring one back either. --deep-link claude-code/permissions/properties had been documented as working the whole time. Settings pages, Effective Settings and Backup / Restore now all round-trip their tab, and a tab contributed by the per-group customizer addresses exactly like a built-in one — --deep-link claude-code/hooks/hooks.flow opens the Hooks flow diagram.

• Plugin artifacts forgot where you were. A remembered position is qualified as name@source, and a plugin's source is itself a path — so the saved path split into too many segments, was rejected, and every plugin artifact's position was discarded on the next launch. Copy deep link handed out the same unusable string. Path separators inside a source are now written as :; both spellings are accepted if you type one by hand.

• Screen readers announced every tab in the app by its class name. Tabs are focusable and selectable, so a screen-reader user lands on one — but unnamed, UI Automation falls back to the bound item's type. Settings pages announced their tabs as Bennewitz.Ninja.ClaudeForge.ViewModels.GroupTab, and Agents & Skills announced its own as Avalonia.Controls.ScrollViewer. Every tab now announces its visible name.

Added

• docs/YAML-FRONT-MATTER.md — a reference for which YAML tokens the editor supports, which round-trip verbatim by design (nested mappings, anchors, aliases, tags), and why implicit type coercion is deliberately not applied so name: yes never becomes name: true in a file you only opened.

• New deep-link targets, documented in the README: claude-code/hooks/hooks.flow, effective-settings/json, backup-restore/restore. Every settings page accepts properties, effective and json.

Changed

• The accessibility guard now covers tabs. TabItem was absent from its list of interactive controls, so that entire class of control had gone unexamined in every file since the test was written. Adding it is what stops the tab-naming problem returning.

• An unrecognised tab id in a deep link is now reported as a miss rather than being applied silently and ignored.