Which agent skills should you stack to build agents and MCP servers?
Reviewed . Scores and counts come from the directory snapshot of .
Short answer: mcp-builder from anthropics/skills (89/100), with about 136,000 installs recorded by skills.sh, guides MCP servers in Python with FastMCP or TypeScript with the MCP SDK. prompt-optimizer from getsentry/skills (81/100) writes the agent, system and developer prompts and the evals, and context-optimization from muratcankoylan/Agent-Skills-for-Context-Engineering (86/100) keeps the context budget under control. For multi-agent work, slicing-code-context, grok-build and executing-plans cover delegation, and aws-agentic-ai covers deployment on Bedrock AgentCore.
The stack
Seven skills, in score order. Scores come from the directory snapshot of October 22, 2026, and install counts are the figures skills.sh published on November 2, 2026.
| Skill | Repository | Score | What it does |
|---|---|---|---|
| mcp-builder | anthropics/skills | 89/100 | Guides building MCP servers that let LLMs use external services through tools, in Python with FastMCP or TypeScript with the MCP SDK. |
| slicing-code-context | trailofbits/skills | 89/100 | Selects bounded, graph-informed source slices with Trailmark and delegates focused code analysis or patch-proposal work to a smaller subagent. |
| grok-build | sanjay3290/ai-skills | 87/100 | Delegates well-specified implementation tasks to xAI's Grok Build CLI running headlessly while the assistant writes task specs, reviews every diff and owns the result. |
| context-optimization | muratcankoylan/Agent-Skills-for-Context-Engineering | 86/100 | Improves context efficiency through context budgeting, observation masking, KV-cache strategy, partitioning and token-cost reduction. |
| executing-plans | obra/superpowers | 83/100 | Guides executing an implementation plan inline in the current session as the implementer, when inline execution is chosen or no subagent tool is available. |
| aws-agentic-ai | zxkane/aws-skills | 83/100 | Guides deploying and managing AI agents on AWS Bedrock AgentCore, including MCP tool integration and governance workflows. |
| prompt-optimizer | getsentry/skills | 81/100 | Creates and refines agent, system and developer prompts and templates, ports prompts between OpenAI, Claude and Gemini, and builds prompt evals. |
How to use it
- Write the prompt and the eval first with prompt-optimizer. It creates and refines agent, system and developer prompts and reusable templates, ports prompts between OpenAI, Claude and Gemini, and builds prompt evals. An agent without an eval is a demo.
- Give the agent access with mcp-builder. It guides an MCP server in Python with FastMCP or TypeScript with the MCP SDK. Check the reference servers before you write one; a tool that already exists is a tool you do not maintain. The Claude Code MCP docs cover wiring it in.
- Budget the context with context-optimization: context budgeting, observation masking, KV-cache strategy, partitioning and token-cost reduction. Do this before you add a second agent, because two agents double a context problem.
- Run the plan with executing-plans when you are the implementer and no subagent tool is available. It executes an implementation plan inline in the current session.
- Delegate, but keep the review. slicing-code-context selects bounded, graph-informed source slices with Trailmark and hands focused analysis or patch-proposal work to a smaller subagent, which sees a slice, not the repository. grok-build sends well-specified implementation to the Grok Build CLI running headlessly while the assistant writes the spec, reviews every diff and owns the result.
- Deploy with aws-agentic-ai if you are on AWS. It covers Bedrock AgentCore Gateway, Runtime, Memory, Identity, Observability, Evaluations, MCP tool integration and governance workflows.
Install the stack
npx skills add https://github.com/anthropics/skills/tree/main/skills/mcp-builder
npx skills add https://github.com/trailofbits/skills/tree/main/plugins/trailmark/skills/slicing-code-context
npx skills add https://github.com/sanjay3290/ai-skills/tree/main/skills/grok-build
npx skills add https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering/tree/main/skills/context-optimization
npx skills add https://github.com/obra/superpowers/tree/main/skills/executing-plans
npx skills add https://github.com/zxkane/aws-skills/tree/main/plugins/aws-agentic-ai/skills/aws-agentic-ai
npx skills add https://github.com/getsentry/skills/tree/main/skills/prompt-optimizerThe skills CLI works for Claude Code, Codex, Cursor and other agents, and each skill's page in the directory lists the copy-by-hand routes if you would rather not run npx.
What to check before you install
- Licenses: context-optimization, executing-plans and aws-agentic-ai are MIT; mcp-builder, grok-build and prompt-optimizer are Apache-2.0; slicing-code-context is CC-BY-SA-4.0, a share-alike license.
- Services: grok-build needs xAI's Grok Build CLI, aws-agentic-ai needs an AWS account with Bedrock AgentCore, and slicing-code-context uses Trailmark. prompt-optimizer ports prompts between OpenAI, Claude and Gemini, so expect keys for whichever you test against.
- Read each SKILL.md before you enable it. The directory has not security scanned any skill, and a skill that delegates work to another agent deserves a slower read than most.
FAQ
Do I need an MCP server for every tool?
No. A skill adds method and an MCP server adds access. If the agent can do the job with files and a shell, a skill is enough. Build a server when the job needs a system the agent cannot otherwise reach, and start from the reference servers before writing your own.
Do these skills work in Codex and Cursor?
Yes. All seven use the Agent Skills format, so the same folder works in Claude Code, Codex, Cursor and any other agent that reads a SKILL.md. The skills CLI installs for each of them, and every skill page in the directory shows the commands.
How do I stop a multi-agent setup from eating the context budget?
Three of the picks exist for that: context-optimization budgets and masks, slicing-code-context hands a subagent a bounded slice, and grok-build moves implementation out of the main session entirely. Measure the token cost before and after each one; the summaries describe what the skills do, not what they will save you. executing-plans, at about 250,000 installs recorded by skills.sh, is the most installed pick here.