Agent Tool 0.7.1

Changed

• The dashboard puts the tool list first: the URL field opens from a + button in the header, Check for updates moves to the header, and the update count becomes a filter toggle next to the scope switch.
• A clipboard URL now fills the URL field instead of showing a separate card.
• Rows flagged by Environment diagnostics show a warning or broken mark.
• Detected tools in a URL analysis use the same row layout as the tool list.
• Dashboard text follows the editor font size.
• The dashboard takes the look of an agent-colored panel: each agent has its own color, the selected agent's list sits in a section with its color stripe, name, scope switch and update filter; statuses such as Updates available and Running appear as filled pills, and tool-kind icons use colors that do not overlap the agent colors.

Fixed

• Background dashboard updates now preserve unchanged tool lists and open details, avoiding repeated entrance animations and flicker.
• Dashboard section titles, the "more" toggle, and the project fallback name are now localized.
• Agent tab counts on the dashboard now match the selected scope.

Agent Tool 0.7.0

Changed

• Each tool row now shows the first line of its description, so the list says what a tool does without opening it.
• The detail panel now explains how to call the tool by name and agent, folds a long description to three lines behind Show more, and offers Open the file for a Skill, Subagent, or Rule. Tools without a description now say what their kind is instead of only "No description was provided."; a Plugin's version moved to its own row.
• The browser extension now shows the description of a Skill or Subagent on the install card, read from the first 4 KB of its source file before you install it.
• The browser extension can now scan unsupported pages with local detection even when AI-assisted detection is disabled or has no Jev API key. It asks for Jev setup only when local evidence cannot decide.
• When automatic or manual detection finds several direct Skill or Subagent links, the popup now lets you choose one and verifies only that choice before opening the install flow.
• Unsupported-page scanning now reads installation commands one line at a time. It recognizes a single --skill argument from npx, bunx, or pnpm dlx, verifies that Skill through GitHub, and never executes the command.
• A stale detection request can no longer replace the current tab's candidate after a navigation, dismissal, or tab close.
• Automatic detection now retries twice within five seconds when a permitted page adds its tool links after initial rendering.
• Large HTTP responses and JSON-LD blocks are now cut off while they are read rather than after, and a registry that would exceed its 2 MB limit is rejected before it is saved instead of being written back unreadable.
• Each site listed under Auto-detect in the browser extension's settings is now a link that opens that site in a new tab.
• Simplified the browser extension popup's visual hierarchy to use one accent color, quieter surfaces, and sentence-case labels.
• The Agent Tool panel in the editor now moves the way the browser extension does: hover and press states settle instead of snapping, a banner, a preview card, and an opened tool detail rise into place, and a collapsible section rotates one caret instead of swapping glyphs. All of it stops under Reduce motion.
• The panel no longer rebuilds its list every three seconds while watching MCP servers; it redraws only when a server starts or stops.
• Shortened the browser extension's detection animation so a detected tool settles in about half the time.

Fixed

• Skills that Claude Code keeps in ~/.claude/skills/synced are now listed as bundled instead of being reported as a directory with no SKILL.md. Agent Tool reads them but never writes there, because Claude Code recreates that directory on each sync.
• Skills that live only in agent-bundled roots are no longer diagnosed. Two agents shipping a skill of the same name is normal, and you cannot remove either copy.
• An archive whose entry declares an unreadable size is now rejected with the same message as any other malformed archive, instead of failing with an internal error.
• A long install no longer loses the write lock to another window. The lock now records the window that holds it, so copying a large Skill is no longer mistaken for a crashed process.
• A registry.json written by a newer version of Agent Tool is now always refused, instead of being read and saved back with its unknown fields dropped.
• Enable for this site now appears for every confirmed result — a single tool, a folder listing, or a candidate list — instead of only the single-tool card. A site whose pages resolve to a folder listing, such as Supabase's skills guide, could not be enabled again once it had been removed.
• Removing a site under Settings → Automatic detection now also drops what automatic detection already showed for that tab, so the popup no longer answers with the removed site's previous result.
• The browser extension now resolves the active browser tab from a popup or service worker correctly, so enabling automatic detection opens the site-permission prompt and rescans the current page after approval.
• A multi-candidate unsupported-page scan no longer shows an empty install card or divider above its candidate list.
• A page that does not offer any installable tool is no longer reported as an unsupported MCP server or plugin. The assisted scan now checks whether the page offers a tool before it looks at the kind.
• When Jev is temporarily overloaded, the assisted scan now says so and asks you to try again later, instead of reporting that the page could not be analyzed.
• Adding, renaming, or deleting a file under a project's .claude/rules or .cursor/rules now refreshes the view.

Agent Tool 0.6.0

Changed

• Repository archives are fetched as tar.gz and unpacked through the same reader the browser extension already used, so the extension now ships with no runtime dependencies and one archive format instead of two.
• The one-time restore of tools disabled by version 0.4.0 has been removed. If you still have such tools, install 0.5.0 once before upgrading, or move them out of the extension's disabled-skills / disabled-agents folders by hand.
• The detection card in the browser extension keeps its entrance animation but no longer draws the sparkle marks beside it.

Fixed

• A GitHub /blob/ URL that points at a directory (GitHub itself redirects these to the /tree/ view) is now read as a directory, so Skills published that way resolve to the right name and path instead of failing verification.

Added

• The browser extension can now detect Skills and Subagents on an unsupported site automatically, one site at a time. After it actually confirms a tool on a page, the popup offers Enable for this site; allowing it grants access to that single origin — never subdomains, never all sites — and Agent Tool then checks pages on that site as you browse. Automatic detection resolves what it can on its own and verifies against GitHub; a page it cannot resolve goes through the same single Jev question the manual scan asks, so it needs AI-assisted detection enabled and is capped per hour (only a counter is stored for that, never the sites you visited). Sites you allowed are listed under Settings → Automatic detection, read back from the browser's own permission API, and removing one stops detection immediately. Navigations on sites you have not allowed never reach the extension.
• Beta: the browser extension can optionally use a user-supplied Jev API key to manually inspect an unsupported page for a GitHub-backed Skill or Subagent. It uses temporary activeTab access, sends only minimized candidate evidence to TypeSafe with key-like values redacted, and still requires the existing GitHub verification before showing an install candidate. Which tool to install is decided locally by matching the page's own URL against the links it carries — Jev is only asked whether the page ships a tool at all, and only when no direct link is found. A page that lists several tools tells you to open the one you want instead of guessing, and a #fragment in the URL scopes the scan to the entry you are reading.

Agent Tool 0.5.0

Changed

• Tool disabling has been removed. Tools disabled by an earlier version are restored when the IDE Dashboard next opens in a writable local window.

Fixed

• Installed managed Skills can now be added to another AI agent from the IDE Dashboard or browser extension without removing the original.
• The browser extension no longer loses the "Installed" state after a browser restart or when the service worker is cold; the popup now trusts the local install record as long as the directory handle is still known and its permission is not denied.

Agent Tool 0.4.0

Added

• The IDE Dashboard header now links to the Agent Tool browser extension in the Chrome Web Store.

Changed

• The IDE Dashboard has a refreshed layout: tools are grouped by kind with counts, agents and scopes are segmented controls (User Global first and selected by default), a banner filters to tools with updates in one click, and Environment and Diagnostics are collapsible footer sections.
• The browser extension popup has a refreshed layout: the Install button leads the detected card, and the settings screen is split into Destinations, Behavior, and Palette with a toggle for auto-open and a count on the installed-items list.
• Opening the browser extension popup on a Tool page you installed from the same GitHub repository now shows an "Installed" tag instead of the Install button. It appears only when the folder is confirmed through the granted folder access, and never sets a badge or opens the popup automatically.
• Tool pages that only share a name with an installed Skill or Subagent from another source now show the install detection instead of staying silent. The overwrite confirmation still protects the existing folder.

Fixed

• The browser extension toolbar badge now uses white text, which some Chrome builds rendered in grey.

Agent Tool 0.3.0

Added

• The Dashboard now shows read-only environment diagnostics for broken links, missing SKILL.md files, independent same-name skill copies, and MCP launch commands that are absent from PATH. It never starts configured MCP commands or records diagnostic history.
• The Dashboard now also warns when Claude registers an MCP server with the same name at user and project scope, without guessing which registration takes precedence.
• The Environment section now separates each agent's supported formats, installed formats, and CLI detection state. Cursor is shown as configuration-detected because it has no CLI interface.
• Update previews now summarize added, removed, and changed files, flag manifest/configuration changes, and say when the displayed diff is incomplete.
• Managed, enabled Skills can now be copied between the user profile and the current project. The confirmation shows both paths, never overwrites an existing destination, and leaves the source unchanged.
• Environment and environment-diagnostics details now sit at the bottom of the Dashboard, are left-aligned and collapsed by default, and can be expanded without saving a UI preference.
• The Environment details now show the absolute path of each detected agent CLI.
• Rules are now listed in the Dashboard for both user scope (~/.claude/rules/*.md, ~/.cursor/rules/*.mdc) and project scope (<project>/.claude/rules/*.md, <project>/.cursor/rules/*.mdc), including rules kept in subdirectories. You can delete rules that you or your team placed there; the extension does not add rules from a URL, and enable/disable are not offered because rules load unconditionally or based on their own paths: frontmatter. Codex AGENTS.md and Gemini GEMINI.md are single-file formats and out of scope.

Changed

• Documented Remote (SSH / Dev Container / Codespaces) behavior as read-only listing with all write operations rejected, matching what the extension has always done. Previous wording implied even reading was blocked.
• Added an opt-in Playwright end-to-end script (npm run test:browser) that loads the packaged browser extension into Chromium and asserts the toolbar badge fills in for each supported catalog. Excluded from CI so real-site or catalog outages never break npm test.
• The browser extension now detects catalog skills that live under .agent-skills/<name>/SKILL.md, checked only when the canonical skills/<name>/SKILL.md returns 404. This lets repositories that keep 100+ skills in a hidden directory (measured: akillness/jeo-skills, 140 MB archive) install a single skill without hitting the archive size limit.
• The browser extension now shows a "!" toolbar badge (with a tooltip explaining the cause) when GitHub's unauthenticated rate limit stops it from listing skills at a directory URL. Previously the badge stayed empty, indistinguishable from a directory that has no skills.

Agent Tool 0.2.1

Added

• MCP servers pinned to @latest are now flagged in the Dashboard, because their contents change on every launch and no version is recorded to update from.
• Subagents are detected whether or not their front matter lists tools:, so the IDE extension now installs the many subagents that omit it.

Fixed

• The browser extension now keeps detected candidates and their badge when automatic popup opening is off, so opening the toolbar action still shows the current candidate.
• Update previews now obey the same untrusted-workspace and Remote-environment guard as other IDE operations.
• Tools installed by the browser extension can now be removed, disabled, re-enabled, and updated from the Dashboard. They were listed as managed, but every one of those actions looked for the files in the IDE extension's own store and reported that the tool was not found; applying an update wrote the new version to that store and left a second copy behind. The registry now records where the files actually are.
• Pinning a tool is no longer silently cleared when the browser extension installs the same tool again.
• Skills, subagents, and plugins can be installed from repositories whose default branch is not main. The download and the update check both asked for main by name, so those repositories returned "not found"; they now ask for the default branch itself.
• MCP arguments containing spaces reach the agent's CLI intact on Windows. Header and environment values such as -H "Authorization: Bearer …" were split into separate arguments, so those servers were registered incorrectly; & and | in a server URL are also accepted now.
• The Command Palette no longer lists six Agent Tool commands that did nothing when run from there, because they act on a tool selected in the Dashboard. They are still available from a tool's actions menu.
• The browser extension declares the popup's language, so screen readers no longer read a Japanese interface with an English voice, and declares its minimum Chrome version.
• Clicking the browser extension's badge after its background worker has been shut down clears the badge instead of opening an empty popup.

Changed

• A workspace folder name containing HTML characters no longer breaks the Dashboard's filter buttons.
• The first update check after installing this version re-reads every source, because sources without an explicit branch are now recorded under the default branch rather than under main.

Agent Tool 0.2.0

Added

• Tools installed by the browser extension are picked up by the IDE extension on its next scan, so they can be removed, disabled, and updated from the Dashboard like anything else.
• Registry entries whose files are gone are dropped on the next scan, so the list no longer shows tools that were deleted outside Agent Tool.
• A browser extension for Chrome, Edge, and Brave that spots skills and subagents while you browse GitHub, skills.sh, and agentsdirectory.dev, and installs them into Claude Code, Cursor, or Codex. It writes only to folders you pick, and works without the IDE extension installed.
• Opening a folder that holds several skills — github.com/<owner>/<repo>/tree/<branch>/skills, or a repository page on a catalog — lists them in the browser extension with an install button per row, so they go in one at a time rather than all together. Each one is fetched file by file instead of as an archive, so skills can be installed from repositories whose archive is over the download limit (measured: 104 KB taken from a repository whose archive is 116 MB).
• A Supported sites dialog in the browser extension with links to GitHub and every supported catalog.
• A light / dark / system theme setting in the browser extension's popup.
• A unified release/Ver_<semver> workflow that gives both extensions the same version, packages and checksums their artifacts, and attaches them to one GitHub Release, plus a privacy policy (PRIVACY.md) and a Chrome Web Store listing checklist (docs/browser-store-listing.md).
• Conditional Chrome Web Store submission from that workflow, skipped until all of its required GitHub Secrets are configured; submit the approved zip to Edge Add-ons manually.
• The browser extension's popup explains what happens before it writes: which folder the item goes into, that your browser only reveals the folder's name rather than its full path, that the source resolves to a public GitHub repository, and that unsafe archive paths and links are refused.

Changed

• Installing stops before anything is deleted when the download contains a name no file system can take (Windows device names, <>:"|?*, trailing dots), instead of failing halfway through and leaving neither the old nor the new version.
• Registry entries are kept, not dropped, when a scanned folder cannot be read, so a temporary permission or sync problem no longer loses pinned and disabled state. The unreadable folder is reported instead.
• Installing over an existing Skill now replaces it instead of writing on top of it, so files that only existed in the previous version no longer linger. The old copy is removed only after the new one has been fetched.
• The browser extension's popup uses a palette matching the extension icon, with a few short, prefers-reduced-motion-aware animations (the detected-card entrance, the install spinner, the dialog open).
• The browser extension no longer asks for the unlimitedStorage permission. It keeps only a small amount of data (the folder handles you granted and the list of what it installed), which fits in the ordinary extension storage quota.
• Installing over an existing item stops instead of replacing it when that item is over 64 MB, because the old copy is held in memory so it can be put back if the replacement fails. Nothing is deleted, and the message says to remove the item yourself or pick another destination.

Fixed

• Installing from a catalog link no longer holds the whole repository in memory while it looks for the skill. The conventional skills/<name> location is checked first with a single request, and when it matches only that folder is read (measured: 51.2 MB down to 2.7 KB for one entry). Catalogs whose name differs from the folder fall back to the previous behaviour.
• A skill whose repository archive is over the download limit can now be installed from its own page: when the archive is refused, the files are fetched individually instead (measured: 104 KB taken from a repository whose archive is 116 MB). Repositories whose archive works are unaffected.
• A repository listing that could not be read is no longer reported as "not found". Hitting GitHub's unauthenticated request limit now reads as a fetch failure, which is what it is.
• Choosing a folder in Brave reported what went wrong instead of doing nothing: Brave turns the File System Access API off by default, and the extension now says so and walks through enabling it in brave://flags.
• The "Installed here" list drops entries that were removed from the IDE extension. It could not read the folder to check, so it kept showing them.
• The browser extension's dropdown menus and dialogs stay legible in dark mode; some text could render unreadable against the page's own dark styling before.
• A Skill whose folder contains a symbolic link is no longer offered for install by the IDE extension. Symlinks are never extracted, so installing one produced a copy with files silently missing.
• Pasting a link that is not a skill or subagent says so in the browser extension. The message was wired up but its visibility was inverted, so nothing appeared.
• The browser extension's "paste a URL" label no longer folds the "Supported sites" link into the input's accessible name, the install-succeeded banner is announced to screen readers and keeps focus while visible, and the URL error message is now associated with its field.

Agent Tool 0.1.2

Changed

• Dashboard rows now show an "Updates available" marker, so it is clear which tools the update count refers to.
• Clicking the "Updates available" count filters the list down to the tools that have an update; clicking it again clears the filter.
• The Dashboard shows a loading indicator until the first scan finishes, instead of looking like an empty or failed list.
• The tool list is posted before the agent CLI scan, so the first view no longer waits on spawning login shells.
• MCP status checks no longer start a new round while the previous one is still running.
• Packaging now starts from a clean build directory.

Fixed

• Update previews, configuration files, registry reads, and GitHub API responses are now bounded, so a single large input cannot grow the extension host's memory without limit.

Agent Tool 0.1.1

Ver_0.1.1

docs: Update extension description in README and localization files f…