Plugins for git and project workflows, code audits and release checks. All 15 run in Claude Code, and 9 of them also run in Codex and Cursor. Each plugin adds its own commands, so install only the ones you need.
Coming from
claude-code-plugins-dev? Version 3.0.0 renamed the repository and the marketplace. Your current install keeps working. See Moving from claude-code-plugins-dev to switch to the new name.
Most skills are commands you run by name, such as /git-init in Claude Code and Cursor. Interactive skills ask for what they need; the rest run directly on your codebase. A few take flags, such as /seo-fix --dry-run and /swift-verify --fix. See each plugin's README.
| Plugin | What it does | Commands | Runs in |
|---|---|---|---|
| ai-accessibility | WCAG audits of code or a live URL, with visual checks through Playwright MCP when it's set up | /accessibility-audit |
Claude Code, Codex, Cursor |
| ai-ado | Azure DevOps work items, hour logging and timesheets through the Azure DevOps MCP server | /ado-init, /ado-create, /ado-log-story-work, /ado-timesheet-report |
Claude Code |
| ai-compliance | License audits of open-source dependencies, plus NOTICE and ATTRIBUTION file generation | /compliance-license-audit, /compliance-notice-generate |
Claude Code, Codex, Cursor |
| ai-git | .gitignore generation, preflight checks (typecheck, lint, format, security and tests), and commit, push and PR creation | /git-init, /git-preflight, /git-commit-push-pr |
Claude Code, Codex, Cursor |
| ai-improve | Audits a codebase across nine categories and writes self-contained implementation plans that another agent or a cheaper model can carry out. It also writes a plan from a description or from another audit's report, hands a plan to an executor in a separate git worktree and reviews the result, and keeps the plans up to date. Its skills never change source code themselves | /improve-audit, /improve-plan, /improve-execute, /improve-reconcile |
Claude Code, Codex, Cursor |
| ai-knowledge | A team knowledge base kept in git as Claude Code rules, so each article loads when Claude works on matching files, every session, or on demand | /kb-init, /kb-migrate, /kb-add, /kb-learn, /kb-discover, /kb-query, /kb-prune |
Claude Code |
| ai-modernize | Technical debt audits for code written with older AI models | /modernize-audit |
Claude Code, Codex, Cursor |
| ai-performance | Performance bottleneck audits with an impact score for each finding, and a GitHub Actions audit that measures where CI time goes (machine wait, setup and work) and gives the file, line and seconds saved for each fix | /performance-audit, /performance-ci |
Claude Code, Codex, Cursor |
| ai-rules | A wizard that adds versioned sections of standing instructions to CLAUDE.md from a shared library (reply format, plain wording, PR and commit hygiene, estimates and more), with modules and settings you pick, and upgrades that keep your edits. It also audits your CLAUDE.md files for more than /doctor prompt-audit checks: it removes rules Claude Code already follows, moves rules into settings or the right file, and trims bloated text |
/rules-setup, /rules-cleanup |
Claude Code |
| ai-security | Security audits with Markdown/JSON reports that track findings across runs, a planner for sandboxed audits on a stronger model, a live-site dependency scan, and settings and supply-chain hardening. It adds to Claude Code's native /security-review |
/security-init, /security-audit, /security-cvp-audit, /security-scan-dependencies, /security-supply-chain |
Claude Code |
| ai-seo | Search audits through three lenses: SEO (deprecated patterns LLMs still generate), AEO (being the direct answer in featured snippets and voice assistants) and GEO (being cited in AI answers). It also crawls the live site to check what search engines get, submits new pages for indexing, applies framework-specific fixes, and writes JSON-LD, FAQ sections, question maps and llms.txt | /seo-audit, /seo-crawl, /seo-fix, /seo-schema, /seo-faq, /seo-questions, /seo-llms-txt, /seo-index |
Claude Code, Codex, Cursor |
| ai-slop | Audits user-facing copy, layouts and metadata for AI-written tells and factual errors, such as a privacy policy that misses the trackers the code loads. It also redesigns one screen that looks AI-generated, rewrites text to remove AI writing patterns, prunes comments that narrate the code, and rebuilds an agent's feature as a simpler second version | /slop-audit, /slop-rewrite, /slop-comments, /slop-rebuild, /slop-design |
Claude Code, Codex, Cursor |
| ai-statusline | Adds progress bars, rate-limit widgets, a month-to-date spend budget for accounts without rate limits, and effort-level and sandbox indicators to Claude Code's native /statusline |
/statusline-wizard, /statusline-edit |
Claude Code |
| ai-swift | Swift/iOS/macOS release checks that catch Xcode Cloud and TestFlight blockers before upload | /swift-preflight, /swift-diagnose, /swift-ci-scaffold, /swift-verify, /swift-concurrency-review |
Claude Code, Codex, Cursor |
| ai-workflow | An unattended run that takes a branch through Claude Code, Codex and Cursor Bugbot review, merge and deploy, or stops when the PR is ready to merge. It also runs review loops with models from several providers, takes over other people's PRs and refreshes stale ones. workflow-pr-comments runs on its own and formats every PR comment the agent posts for you |
/workflow-babysit, /workflow-council, /workflow-takeover, /workflow-refresh |
Claude Code |
The Claude Code-only plugins depend on Claude Code subagents, settings files, rules, the status line or CLAUDE.md.
The audit skills take their settings as arguments, so a scheduled or unattended run doesn't stop to ask. For example, /accessibility-audit 2.2 AA src/, /modernize-audit src --defaults, /seo-audit geo, /improve-audit quick --top 3 or /swift-preflight --check-only. In Claude Code, /security-audit can also split its review between parallel security-auditor subagents.
The audit plugins read your code and report what they find. A few can also check a live URL, and
/performance-cimeasures recent CI runs. Use them to catch problems during development, then confirm with runtime testing and manual review.
Install Claude Code, then add the marketplace and install the plugins you want:
/plugin marketplace add charlesjones-dev/ai-agent-plugins
/plugin install <plugin-name>@ai-agent-plugins
# For example:
/plugin install ai-git@ai-agent-plugins
/plugin install ai-security@ai-agent-plugins
Then run any of the plugin's commands:
/git-init # Generate a .gitignore for the project
/security-init # Add file-access deny rules for secrets
/ado-init # Set up Azure DevOps and its MCP server
codex plugin marketplace add charlesjones-dev/ai-agent-plugins
codex plugin add <plugin-name>@ai-agent-plugins
# For example:
codex plugin add ai-git@ai-agent-plugins
Codex lists only the plugins marked Codex in the table and prefixes each skill with its plugin name, such as ai-git:git-init. Start a new session after installing, then type $ to pick a skill or open /skills. Codex runs a skill only when you pick it, with six exceptions: it can start the slop audit when you ask whether a site or README reads as AI-written, the design review when you ask it to unslop or polish a screen that looks AI-generated, the comment review when you ask it to prune or clean up code comments, the rebuild when you ask it to scrap an implementation and redo it the simple way, the git preflight checks when you ask it to check your code before a commit, and the index submission when you ask it to get pages indexed. Plugins don't load in the Codex IDE extension, so use the Codex CLI or the ChatGPT desktop app.
Cursor installs plugins from team marketplaces on its Teams and Enterprise plans. An admin adds this repository under Dashboard > Plugins & MCPs > Team Marketplaces > Add Marketplace > Import from Repo, using https://github.com/charlesjones-dev/ai-agent-plugins. Cursor lists only the plugins marked Cursor in the table.
To use a plugin on one machine instead, copy its folder to ~/.cursor/plugins/local/, for example plugins/ai-git to ~/.cursor/plugins/local/ai-git, then run Developer: Reload Window. On Teams and Enterprise plans, local plugins load only if an admin turns on Allow Local Plugin Imports, which is off by default on Enterprise. Run a skill with / and its name, such as /git-init.
The plugins that run in Codex and Cursor follow the Agent Plugins 1.0 format, with a plugin.json at the root of each plugin folder. Other agents that support the standard can load them from plugins/<plugin-name>.
The skills are also listed on skills.sh. Its skills CLI copies the SKILL.md files into agents that support the Agent Skills format:
npx skills add charlesjones-dev/ai-agent-plugins # All skills
npx skills add charlesjones-dev/ai-agent-plugins --skill <skill-name> # One skill
# For example:
npx skills add charlesjones-dev/ai-agent-plugins --skill git-init
This path copies only the skill files. It doesn't replace the marketplace install in Claude Code, Codex or Cursor. On those agents, install from the marketplace to get the full plugin. In Claude Code, the full plugin also includes the subagents, settings, status line and knowledge-base tooling that the Claude Code-only plugins need.
Version 3.0.0 renamed the repository from claude-code-plugins-dev to ai-agent-plugins, and the marketplace to match. GitHub redirects the old URLs, so existing installs and links keep working.
Claude Code keeps your plugins under the old marketplace name, so commands that use @ai-agent-plugins won't find them. To switch:
/plugin marketplace remove claude-code-plugins-dev
/plugin marketplace add charlesjones-dev/ai-agent-plugins
/plugin install <plugin-name>@ai-agent-plugins
Removing the old marketplace uninstalls its plugins, so reinstall the ones you use.
Four plugins merged into others. Claude Code switches an existing install's setting over to the plugin that replaced it, but you still need to install that plugin once:
| Old plugin | Now part of | Run once |
|---|---|---|
| ai-writing | ai-slop | /plugin install ai-slop@ai-agent-plugins |
| ai-aeo, ai-geo | ai-seo | /plugin install ai-seo@ai-agent-plugins |
| ai-cvp | ai-security | /plugin install ai-security@ai-agent-plugins |
These commands were renamed or merged into another command:
| Old command | New command |
|---|---|
/writing-humanize |
/slop-rewrite |
/aeo-audit, /geo-audit |
/seo-audit aeo, /seo-audit geo (or /seo-audit for all three lenses) |
/aeo-fix, /geo-fix |
/seo-fix |
/aeo-faq, /aeo-questions, /geo-llms-txt |
/seo-faq, /seo-questions, /seo-llms-txt |
/cvp-defense-audit |
/security-cvp-audit |
/git-commit-push |
/git-commit-push-pr --no-pr |
/workflow-ship |
/git-commit-push-pr --preflight, in ai-git |
/workflow-preflight |
/git-preflight, in ai-git |
/modernize-scan |
/modernize-audit <path> --defaults |
/ado-create-feature, /ado-create-story, /ado-create-task, /ado-work-items |
/ado-create feature, /ado-create story, /ado-create task |
/kb-ingest, /kb-import, /kb-harvest, /kb-absorb |
/kb-add with a file, folder, repo or URL, or /kb-add --absorb |
/kb-search, /kb-list |
/kb-query search <terms>, /kb-query list |
/kb-organize, /kb-remove |
/kb-prune organize, /kb-prune remove <article> |
/kb-upgrade |
/kb-migrate |
ai-knowledge 2.0 also changes where the knowledge base lives. See Knowledge bases from ai-knowledge 1.x.
If you used /workflow-ship or /workflow-preflight, install ai-git. Commands removed without a replacement in this marketplace are listed under Deprecated & Removed.
3.0.0 also removes the accessibility-auditor, modernize-auditor, performance-auditor, security-dependency-scanner and swift-release-auditor subagents. They couldn't load the skills they wrapped. Run the matching skill with arguments instead, such as /performance-audit or /swift-preflight --check-only.
In Codex, run codex plugin marketplace remove claude-code-plugins-dev, then follow the Codex steps.
ai-knowledge 2.0 stores the knowledge base as Claude Code rules instead of a table in CLAUDE.md, so Claude Code loads each article at the right time by itself.
In 1.x, articles lived in docs/kb/, and a table in CLAUDE.md told Claude when to read each one. The table loaded in every session, and an article only reached Claude if Claude read the table, matched a row and chose to open the file. Claude Code loads files in .claude/rules/ on its own:
- Scoped articles live in
.claude/rules/kb/withpathsfrontmatter, such assrc/api/**. Claude Code loads one when Claude reads, writes or edits a matching file, and a session that never touches those files spends no context on it. - Pinned articles also live in
.claude/rules/kb/, withoutpaths, and load at the start of every session, as they did before. - On-demand articles cover topics rather than files, such as release steps or an external system. They live in
.claude/kb/, each with one line in.claude/rules/kb/_index.md, which loads every session, and Claude opens an article when a task matches its line. This is the only part that still depends on Claude choosing to read a file, the way the whole 1.x table did.
_index.md also lists each scoped article with its keywords, so Claude can read one when the topic comes up before it touches a matching file, as a 1.x row's keywords allowed. When a scoped or on-demand article loads, Claude says so in one line, such as 📖 Loaded KB: conventions/api.md.
Because Claude Code reads .claude/rules/ itself, teammates get the knowledge base from git without installing the plugin, and CLAUDE.md no longer carries a table. The plugin is what you use to capture, add, query, prune and migrate articles.
The articles can't stay in docs/kb/ behind short rule files that point to them. Claude Code expands an @ import inside a rule at the start of every session, even when the rule's paths don't match, so every article would load every time.
1.x had /kb-auto to remind you to capture learnings. 2.0 replaces it with auto-capture: when you turn it on, Claude saves corrections, rules and decisions the team should know as they come up, without you running /kb-learn, and tells you in one line what it saved so you can undo it. /kb-init and /kb-migrate turn it on for you unless you pass auto=off. It's set per developer, so it doesn't change anything for teammates, and it works whether Claude Code's auto memory is on or off. Everything it writes shows up in your diff before it reaches anyone.
Codex and Cursor don't load .claude/rules/, so /kb-init and /kb-migrate also add a short section to AGENTS.md that tells those agents how to read the knowledge base (agents=off skips it). They can read it but not add to it: the /kb-* commands and auto-capture run only in Claude Code.
Two things change for you:
- The knowledge base is now in a hidden folder. If your
.gitignoreexcludes.claude/, it won't reach git. Every/kb-*command checks for this and suggests a fix. - 2.0 drops Obsidian support. The
relatedandsourcesfrontmatter and[[wiki-links]]become plain paths in an optional Related section, which Claude can follow.
Run /kb-migrate once in each project that has a 1.x knowledge base, or /kb-migrate --dry-run to preview it first. It sorts each article into scoped, pinned or on-demand from its table row and frontmatter, moves it with git mv so its history follows, keeps each row's keywords in its tags, turns links into plain paths, and removes the CLAUDE.md table, the /kb-auto block and docs/kb/.obsidian/. At the end it turns on auto-capture for you, unless you pass auto=off. It also finds notes that still name renamed or removed commands, such as /writing-humanize or /geo-audit, in the knowledge base, CLAUDE.md, AGENTS.md and docs/, and proposes fixes. Run it again after a later update to catch new ones. The ai-knowledge README has the details.
Plugins and skills that Claude Code now covers natively, or that moved into another plugin, are retired here. Removed plugins and skills live on in git history.
| Item | Status | Superseded by | Details |
|---|---|---|---|
ai-writing (/writing-humanize) |
🗑️ Removed (October 2026, v3.0.0) | ai-slop /slop-rewrite |
Same rewrite, now sharing one signal list with /slop-audit, and it takes pasted text or a file path as arguments |
| ai-aeo, ai-geo | 🗑️ Removed (October 2026, v3.0.0) | ai-seo /seo-audit aeo, /seo-audit geo |
One plugin and one audit for SEO, AEO and GEO, with a score per lens. All checks and fixes carried over |
ai-cvp (/cvp-defense-audit) |
🗑️ Removed (October 2026, v3.0.0) | ai-security /security-cvp-audit |
Same sandboxed audit planner, next to /security-audit |
ai-workflow /workflow-claude-actions |
🗑️ Removed (October 2026, v3.0.0) | Native /install-github-app |
Its review workflow posts inline comments, and both workflows run on Claude Code's default model (Opus 5.5 in October 2026). To refresh older workflows, run it again and pick "Update workflow file with latest version" |
ai-workflow /workflow-principles |
🗑️ Removed (October 2026, v3.0.0) | Native /init |
/init writes CLAUDE.md from the project's own build commands and conventions |
ai-workflow /workflow-rules |
🗑️ Removed (November 2026, v3.6.0) | ai-rules /rules-setup |
Its five PR and commit rules are one section of a larger library with versions, modules and upgrades. /rules-setup adopts rules installed by /workflow-rules and keeps your edits. Install ai-rules once |
ai-git /git-pr-codex-loop |
🗑️ Removed (October 2026, v3.0.0) | Native /code-review --fix and /autofix-pr |
Both loop on review comments and CI failures. /workflow-babysit still triages Codex Code Review findings as part of its run |
ai-knowledge /kb-auto |
🗑️ Removed (October 2026, v3.0.0) | ai-knowledge auto-capture | /kb-init and /kb-migrate turn it on unless you pass auto=off. Claude saves team rules, corrections and decisions to the knowledge base as they come up, and reports each save in one line that offers an undo. It's set per developer and works with Claude Code's auto memory on or off |
ai-knowledge /kb-load |
🗑️ Removed (October 2026, v3.0.0) | Native @-mentions | Mention the article's file, such as @.claude/kb/deploys.md. Scoped and pinned articles now load by themselves as rules |
ai-learn (/learn, /learn-review) |
🗑️ Removed (July 2026) | Native Learning output style (/config → "Output style" → "Learning") |
Built-in collaborative mentor mode with TODO(human) markers; an Explanatory style also ships natively |
ai-workflow /workflow-plan-phases |
🗑️ Removed (July 2026, v2.0.0) | Native Dynamic Workflows | A built-in runtime runs subagents with a token budget for each, so you don't size phases at 30–50k tokens by hand |
ai-workflow /workflow-implement-phases |
🗑️ Removed (July 2026, v2.0.0) | Native Dynamic Workflows | The runtime coordinates the subagents, so you don't fan out Task() calls or track phases by hand |
Several ai-workflow skills, the PR guidance in ai-git's /git-commit-push-pr, and one PR rule in ai-rules are adapted from tstack, Theo Browne's agent skills for Claude Code and Codex, some of which use T3 Code features. tstack is MIT licensed. Each plugin that uses it ships a THIRD-PARTY-NOTICES.md with the license and a list of the adapted files: ai-workflow, ai-git and ai-rules.
ai-improve is adapted from improve, shadcn's MIT-licensed agent skill that audits a codebase and writes implementation plans for other agents. The plugin splits that one skill into four commands and rewrites its text. Its THIRD-PARTY-NOTICES.md has the license and a list of the adapted files.
ai-performance's /performance-ci is inspired by a post by Vox (Voxyz_ai on X) with a prompt for a CI audit. It follows the post's approach: split each job's time into queue, setup and work, check nine causes, and report before changing anything. Its text and measurement script are original.
ai-slop's /slop-design is inspired by unslop, Damir Sokolovsky's (@heydamir) CC BY 4.0 skill for redesigning AI-looking web UI. It follows unslop's approach (an audit, then a direction, then ordered passes) and adds iOS and Android guidance. Its text and reference files are original, so it ships no copy of unslop's files.
Bug reports and fixes are welcome. See CONTRIBUTING.md.
MIT License. See LICENSE.
Charles Jones
- Website: charlesjones.dev
- GitHub: @charlesjones-dev
- Claude Code plugins
- Codex plugins
- Cursor plugins
- Agent Plugins specification
- Agent Skills specification