AI Agent Stack Directory: Skills, Plugins, MCP, ACP, Prompts and Memory

An agent stack is easier to understand as six layers:

1. Instructions — project rules such as AGENTS.md, CLAUDE.md or README.md.
2. Skills — reusable methods, normally packaged as a SKILL.md plus references and scripts.
3. Plugins — distributable packages that may add tools, runtime code, hooks or credentials.
4. MCP — the Model Context Protocol for connecting agents to external tools and data.
5. ACP — the Agent Client Protocol for hosting a coding-agent session inside a client or editor.
6. Memory — concise, curated continuity: decisions, preferences and useful project facts.

Where to find skills and plugins

“Ponytail” and “Superpowers” should be listed as community projects, not as built-in OpenClaw features. Names and installation methods can change; link to the canonical repository and show its audit date.

Skills list by job

Coding and product engineering

• Repository exploration and codebase navigation
• Test-driven development, debugging and code review
• Next.js, React, TypeScript and API implementation
• Database schema, migrations and query review
• Documentation and release-note generation
• Security and dependency review

UI/UX and frontend

• UI/UX Pro Max: visual direction, palettes, typography and component patterns
• Frontend design: production-grade page and component implementation
• Web Interface Guidelines: accessibility and interaction review
• Next.js SEO: metadata, sitemap, robots and structured data
• Screenshot-based visual QA and responsive testing

Research and operations

• Source discovery with citations
• Long-document summarisation and extraction
• Spreadsheet and data reconciliation
• SEO content planning and editing
• Calendar, email, issue-tracker and knowledge-base workflows

Crypto and finance

• Read-only market-data retrieval
• Technical-analysis calculation and backtesting
• Filing, earnings and comparable-company analysis
• Wallet, tokenomics and contract-risk research
• Tax export and transaction classification

Treat these as capability categories. A skill should document inputs, outputs, failure modes, tools used and permissions; a list of buzzwords is not a useful skill specification.

Plugin list

Prefer plugins that are official, maintained, versioned and narrow in scope:

• GitHub, GitLab or issue-tracker integration
• Documentation and browser research integration
• Figma or design-file integration
• Database integration using a read-only role
• Spreadsheet and document export
• Calendar, email or Slack integration with draft-only defaults
• OpenClaw gateway plugins for approved services

Do not grant a plugin production secrets, wallet signing, withdrawal access or unrestricted filesystem access by default.

MCP list

Useful MCP categories include:

• Filesystem: selected project directories only
• GitHub: issues, pull requests and repository context
• Browser: isolated research and visual verification
• Docs and knowledge bases: approved spaces only
• Figma: inspect files and design context
• PostgreSQL: read-only replica or sandbox database
• Playwright/browser testing: local test environment
• Search and market-data APIs: read-only, rate-limited access

MCP gives an agent capabilities. It does not decide whether the agent should use them. Add explicit approval around every write, delete, publish, payment and trade action.

ACP list

ACP is a session protocol between a coding agent and a client/editor. It is useful for:

• Starting and resuming agent sessions
• Sending prompts and receiving progress updates
• Permission requests before file operations
• Applying or rejecting file changes
• Cancellation, errors and structured handoffs

ACP is not an MCP replacement and is not a trading protocol. MCP connects tools; ACP helps a client host and control an agent.

Prompt patterns that actually copy and run

Use real fenced code blocks in the source markdown. Never place a fence inside a paragraph or JSX text node.

text
Copy
1You are a coding agent working in a repository.
2First inspect the repository instructions and the relevant SKILL.md.
3State your assumptions and proposed file changes before editing.
4Run the smallest relevant tests after each meaningful change.
5Do not modify secrets, production data, or unrelated files.
6Return: changed files, tests run, failures, and remaining risk.

For UI/UX work:

text
Copy
1Act as a UI/UX engineer. Inspect the existing design system first.
2Build the smallest responsive implementation that satisfies the brief.
3Check keyboard navigation, focus states, contrast, reduced motion,
4mobile layout and empty/error/loading states. Do not invent a new
5visual language unless the brief explicitly asks for one.

For research:

text
Copy
1Use cited primary sources. Label each statement as fact, estimate or opinion.
2Separate evidence from inference, list conflicting evidence and identify
3the next verification step. Do not execute trades, publish content or send
4messages without explicit human approval.

Memory workflow

Start with versioned project documentation. Add embeddings or a memory database only when the document set is large enough that retrieval saves time.

The workflow is:

observe → extract durable facts → review → write concise memory → retrieve only when relevant

Memory should be editable, inspectable and deletable. It is a curated index, not a licence to retain everything.

End-to-end workflow

text
Copy
1discover → inspect → choose one skill → grant least privilege → plan
2→ implement → test → read-only review → human approval → record decision

Suggested first pass:

bash
Copy
1# Discovery only; do not pipe an untrusted installer into a shell.
2git clone --depth 1 https://github.com/obra/superpowers.git /tmp/superpowers-review
3sed -n '1,240p' /tmp/superpowers-review/README.md
4find /tmp/superpowers-review -name SKILL.md -print

The command above only clones and inspects a community repository. Installation is agent-specific and should follow that repository’s current instructions after review.

Markdown/rendering fixes for this article

• Put fenced blocks directly in Markdown, not inside <p>...</p>.
• Do not put <div> or <img> inside a paragraph element.
• Escape literal backticks when they are explanatory text; use a real code block for executable examples.
• Keep audit instructions as normal prose or a blockquote; otherwise a renderer may turn a sentence such as “inspect the repository and SKILL.md before adding anything” into an accidental H1.
• Run a rendered-page smoke test: headings, code-block copy buttons, links, tables, images and mobile overflow.

Final rule

The useful stack is small: skills for repeatable methods, plugins for packaged integrations, MCP for external tools, ACP for portable coding-agent sessions, prompts for task constraints and memory for concise continuity. Install less, inspect more and keep write access behind approval.

Related reading

• July 2026 frontier model comparison
• Kimi K3: architecture, benchmarks and open weights
• Grok 4.5 coding model explained

Related guides

• Kimi K3 Is Live: 2.8T Open Frontier Intelligence
• The Complete Guide to Model Context Protocol (MCP) in 2026: Building the USB-C for AI Agents
• Next.js 16.2: 400% Faster Dev & AI-Native Framework