One command to give every AI coding agent you use the same memory, reviewers, and working rules.
Coding agents start every session from zero. They forget what your team decided last week, skip the conventions in AGENTS.md, guess at library versions, and review code in one generic pass. myagents fixes that across Claude Code, Cursor, and Codex at once:
- Memory that persists. Agents look up project decisions and your preferences before they start, and save new ones when they finish. Memory follows the repository across clones and machines.
- 14 specialist reviewers. Security, performance, architecture, tests, accessibility, database, dead code, and more. Run all of them in parallel, pick one, or point them at a GitHub pull request.
- Rules that stay on. Research before guessing, the simplest change that works, no filler in answers. Hooks load them at the start of every session and repeat them on every prompt, so they hold up in long conversations.
- One install, three tools. The installer finds the tools you have and sets up each one the way that tool expects. Rerun it to update.
On macOS, run this. It installs any missing tools it needs with Homebrew, and installs Homebrew first if you don't have it:
curl -fsSL https://raw.githubusercontent.com/RashadAnsari/myagents/master/install.sh | bashmyagents
✓ Download Saved to ~/.myagents (version 65f708b).
✓ Requirements All installed.
✓ Apps Found Claude Code, Cursor, Codex.
albino
✓ Claude Code Installed. Restart open sessions to load it.
✓ Cursor Installed. Run 'Developer: Reload Window' to load it.
✓ Codex Installed. Restart Codex, then run /hooks once to trust its hooks.
archify
✓ Claude Code Installed. Restart open sessions to load it.
✓ Cursor Installed. Start a new chat to load it.
✓ Codex Installed. Restart Codex to load it.
context7
✓ Claude Code Installed. Restart open sessions, then run /mcp once to sign in to Context7.
✓ Cursor Installed. Sign in to Context7 when Cursor asks, or from Cursor Settings > MCP.
✓ Codex Installed. Run 'codex mcp login context7' once to sign in.
emil-skills
✓ Claude Code Installed. Restart open sessions to load it.
✓ Cursor Installed. Start a new chat to load it.
✓ Codex Installed. Restart Codex to load it.
ponytail
✓ Claude Code Installed. Restart open sessions to load it.
✓ Codex Installed. Restart Codex, then run /hooks once to trust its hooks.
✓ Cursor Installed. Start a new chat to load it.
security-audit
✓ Claude Code Installed. Restart open sessions to load it.
✓ Cursor Installed. Start a new chat to load it.
✓ Codex Installed. Restart Codex to load it.
ui-ux-pro-max
✓ Claude Code Installed. Restart open sessions to load it.
✓ Cursor Installed. Start a new chat to load it.
✓ Codex Installed. Restart Codex to load it.
Summary
✓ albino Ready.
✓ archify Ready.
✓ context7 Ready.
✓ emil-skills Ready.
✓ ponytail Ready.
✓ security-audit Ready.
✓ ui-ux-pro-max Ready.
Then try it in any project:
/reviewcrew # every reviewer in parallel, report in REVIEW_REPORT.md
/audit security src/api # one reviewer, one folder
/pr-review https://github.com/org/repo/pull/42
/remember we deploy from the release branch, never from master
To update, run the same install command again.
The diagram is generated with archify from docs/architecture.json. Run make diagram to render and export it again.
- Architecture
- Plugins: albino, archify, context7, emil-skills, ponytail, security-audit, ui-ux-pro-max
- Platform support
- Agent memory
- Installing, updating, and uninstalling
- Contributing
- License
The main plugin: review agents, skills, MCP servers, hooks, and rules.
Skills you run by name. They replaced the old slash commands, which all three tools are moving away from.
| Skill | What it does |
|---|---|
/reviewcrew |
Runs every review agent in parallel and writes a consolidated REVIEW_REPORT.md |
/reportloop |
Walks through REVIEW_REPORT.md one issue at a time, by severity: explains it, then fixes or skips it |
/audit [reviewer] [path] |
Runs one reviewer of your choice, optionally on one path, and prints its findings |
/pr-review <url> |
Reviews a GitHub pull request: picks the reviewers that fit the changed files, loads memory, checks ripple effects, lets you choose which findings to post, and submits the review with gh |
/commit |
Stages all changes and writes a commit message in the project's existing style |
/commit-pr |
Squashes the branch into one commit and opens or updates its pull request, creating a branch first when you are on main or master |
/remember <what> |
Saves something to project or user memory, or this session's learnings when given nothing |
/forget <what> |
Finds and removes matching entries from project and user memory |
/enrich |
Mines the 100 most recently merged pull requests for decisions, conventions, and gotchas, and saves them to project memory |
/statusline |
Sets the Claude Code status line: folder, git branch and dirty state, model, context usage, and time |
Run one with /<name> in Claude Code (or /albino:<name> when another plugin uses the same name) and Cursor, and with $<name> or /skills in Codex. Claude Code fills in the arguments after the name; Cursor and Codex do not pass arguments to skills, so give the subject in the same message instead. commit, commit-pr, pr-review, forget, enrich, reviewcrew, and statusline run only when you type them, never on the agent's own initiative.
Used by /reviewcrew, /audit, and /pr-review, or on their own.
| Agent | Looks for |
|---|---|
security-reviewer |
Injection, XSS, auth flaws, SSRF, JWT misuse, and supply chain risk |
code-reviewer |
Correctness, anti-patterns, concurrency, and memory bugs |
architecture-reviewer |
Structure, coupling, cohesion, SOLID, and duplication |
performance-reviewer |
Bottlenecks, complexity, slow queries, caching, and scalability |
test-reviewer |
Coverage gaps, weak assertions, flaky tests, and over-mocking |
logging-reviewer |
Logging and monitoring gaps, audit trails, and sensitive data in logs |
dependency-reviewer |
Vulnerable, outdated, and unused packages |
docs-reviewer |
Documentation that is wrong, missing, or stale |
agents-md-reviewer |
Code that breaks the rules in AGENTS.md |
accessibility-reviewer |
WCAG, ARIA, keyboard navigation, and screen reader support |
api-design-reviewer |
REST and GraphQL naming, HTTP semantics, versioning, error shape, and compatibility |
database-reviewer |
Schema design, migration safety, indexes, constraints, and query patterns |
i18n-reviewer |
Hardcoded strings, date and number formatting, plurals, RTL, and locale handling |
unused-code-reviewer |
Dead code, tests, CI jobs, infrastructure, config, docs, and assets across the whole repository |
Three skills are always on. The session-start hook tells the agent to load each one in full before its first answer.
| Skill | What it enforces |
|---|---|
dev-conventions |
Think before coding, the simplest change that works, no duplication, surgical edits, and checkable success criteria |
research-first |
Read the docs and source before answering, and use the latest stable version of every dependency |
agent-memory |
Search memory before non-trivial work, and save durable learnings after it, telling you each one in the chat |
The rest load when a task calls for them:
| Skill | Use it for |
|---|---|
frontend-design |
Distinctive, production-grade interfaces that avoid the generic AI look |
humanizer |
Removing the signs of AI-generated writing from text |
markitdown |
Converting PDFs, Office files, images, audio, and URLs to Markdown |
plugin-authoring |
Writing skills and agents that trigger reliably and use few tokens |
eu-ai-act |
EU AI Act risk tiers, obligations, deadlines, and penalties |
| Server | What it adds |
|---|---|
agent-memory |
Project and user memory with semantic search. Rejects vague, duplicate, and secret-containing entries. See Agent memory |
playwright |
Browser automation: navigate, click, fill forms, take screenshots, and inspect the page. Set PLAYWRIGHT_MCP_EXTENSION_TOKEN to drive your real browser through the Playwright Chrome extension |
markitdown |
Converts PDF, Word, PowerPoint, Excel, HTML, images (OCR), audio (transcription), EPUB, and YouTube URLs to Markdown |
| Hook | Runs | What it does |
|---|---|---|
session-start |
Start of every session | Loads the mandatory skills, the memory rules, and the communication rules: answer first, be precise, no filler |
user-prompt-submit |
Every prompt | Repeats a one-line reminder of those rules, since session-start context fades in long conversations |
stop |
End of every turn | Asks the agent whether it learned anything durable and, if so, to save it and tell you what it saved and why |
Instructions every agent follows on every task. Each rule is one plugins/albino/rules/<name>.mdc file.
| Rule | What it says |
|---|---|
single-commit-pr |
A pull request the agent opens has exactly one commit: it squashes the branch before opening it, then amends and force-pushes with --force-with-lease for every later change. It never rewrites a branch someone else pushed to |
Cursor loads the rules from the plugin and lists them under it. Claude Code and Codex cannot, so the installer copies them into ~/.claude/rules/albino-<name>.md, whose prefix marks them as albino's, and into a block between <!-- albino:rules:begin --> and <!-- albino:rules:end --> markers in ~/.codex/AGENTS.md. Edit a rule and it takes effect after the next install: restart Claude Code and Codex, and reload the Cursor window. Reinstalling rewrites only those files and that block, and uninstalling removes them; your own rules and the rest of AGENTS.md are left alone.
Archify by tt-a1i turns a description, a repository, or pasted Mermaid into an interactive architecture, workflow, sequence, data-flow, or state diagram. The result is one self-contained HTML file with dark and light themes, node details, and path tracing, which you can open in any browser or send to someone, and export as PNG, SVG, or video. Ask for a diagram ("map this repository's runtime architecture", "show the checkout flow") and it activates on its own; for a repository, it can link each node to the source files behind it.
When it delivers a diagram it may check for a newer release to show a reminder, and never updates itself; set ARCHIFY_UPDATE_CHECK_DISABLED=1 to turn that check off.
Context7 by Upstash gives agents current, version-specific documentation for libraries and frameworks, pulled from their source repositories, so they stop guessing at APIs that changed after their training data. It pairs with albino's research-first skill, which tells agents to check the docs before answering; Context7 is where they read them. Ask a library question and it resolves the library, then fetches the relevant docs.
Sign in once per tool, so no API key is stored on disk: run /mcp in Claude Code, sign in when Cursor asks, or run codex mcp login context7.
Emil Kowalski's skills teach agents the small interface details they tend to get wrong: the right easing and duration for an animation, a soft shadow instead of a hard border, motion that does not drop frames. They draw on his work at Vercel and Linear. The set covers building, reviewing, and auditing animations (animate, animate-expo, review-animations, improve-animations, find-animation-opportunities, animation-performance, animation-vocabulary), general design (emil-design-eng, apple-design, mobile-native, break-ui, prototype, pick-ui-library), Swift (write-swift), and his Sonner toast library (ask-sonner). They complement albino's frontend-design skill, which sets a visual direction, and ui-ux-pro-max, which supplies styles and palettes; these focus on motion and polish.
ponytail by Dietrich Gebert is a third-party plugin that pushes agents toward the shortest solution that works: YAGNI, standard library first, no abstractions nobody asked for.
security-audit by Cloudflare is the skill that seeded Cloudflare's vulnerability discovery harness. Ask for a security audit and it runs isolated agents through reconnaissance, coverage-led hunting, and validation, where a fresh agent tries to disprove every finding, then writes a report with confirmed findings and open leads. Reports default to ~/security-audit-skill/<repo-name>/. It complements albino's security-reviewer, which gives a quicker single-pass review.
Building or running the audited code needs an OS-level sandbox; without one, the skill reports those leads as unconfirmed instead of executing the target.
UI UX Pro Max by nextlevelbuilder gives agents a searchable local design database: 79 styles, 192 product palettes, 74 font pairings, 119 UX guidelines, 25 chart types, and guidance for 22 stacks, from React and SwiftUI to Flutter and Tailwind. It activates on its own when you build, review, or fix an interface, and can generate a design system for your product. It overlaps with albino's frontend-design skill, which sets a visual direction; this one supplies the concrete styles, palettes, and stack rules.
It ships seven skills: ui-ux-pro-max, design, design-system, ui-styling (shadcn/ui and Tailwind), brand, banner-design, and slides.
| Claude Code | Cursor | Codex | |
|---|---|---|---|
| albino review agents | ✓ | ✓ | ✓ |
| albino skills | ✓ | ✓ | ✓ |
| albino hooks | ✓ | ✓ | ✓ after /hooks |
| albino MCP servers | ✓ | ✓ | ✓ |
| albino rules | ✓ | ✓ | ✓ |
| archify | ✓ | ✓ | ✓ |
| context7 | ✓ | ✓ | ✓ |
| emil-skills | ✓ | ✓ | ✓ |
| ponytail | ✓ | ✓ | ✓ after /hooks |
| security-audit | ✓ | ✓ | ✓ |
| ui-ux-pro-max | ✓ | ✓ | ✓ |
Memory lives in one SQLite database at ~/.myagents/agent-memory/memory.sqlite, outside every repository. Set AGENT_MEMORY_DIR to move it.
- Project memory is tied to the repository's
originremote, so it follows the repo across clones and machines. Repositories without a remote fall back to their local path. - User memory holds your preferences and applies to every project.
project_searchwithall_projects: truesearches every project at once, and each result says which repository it came from.
Uninstalling keeps your memory unless you pass --purge.
Installing. Run the command from Quick start. The installer downloads this repository to ~/.myagents, installs any missing tools from requirements.list with Homebrew (installing Homebrew first if needed, which may ask for your password), then sets up every plugin in each of Claude Code, Cursor, and Codex that it finds. Tools you already have are left alone. git and python3 come with the Xcode Command Line Tools (xcode-select --install).
Each line of the output says what to do next, such as restarting a tool, signing in to Context7, or running /hooks once in Codex to trust the hooks.
Updating. Run the install command again. Every step is safe to repeat: a rerun updates what is there and never duplicates settings, hooks, or links. ~/.myagents belongs to the installer, so a rerun resets it to the latest version and discards local edits to tracked files. Untracked files, such as the memory database, are kept.
Uninstalling. This undoes every install step on every tool, including settings, hooks, and MCP servers, then deletes ~/.myagents. Your own settings, hooks, skills, agents, and rules are left alone, including the text in ~/.codex/AGENTS.md outside the albino block, and so is your agent memory. Homebrew and the tools from requirements.list stay installed, since other software may use them; remove them with brew uninstall if you no longer need them.
curl -fsSL https://raw.githubusercontent.com/RashadAnsari/myagents/master/uninstall.sh | bashAdd --purge to delete your agent memory too:
curl -fsSL https://raw.githubusercontent.com/RashadAnsari/myagents/master/uninstall.sh | bash -s -- --purgeRerunning the uninstaller is safe: steps with nothing left to undo report "Already removed." If a step fails, ~/.myagents is kept so you can fix the problem and rerun.
Repository rules for contributors and agents, including how to add a plugin, are in AGENTS.md. Run make local before opening a pull request.
MIT © Rashad Ansari