Last updated: October 3, 2026
Agent Skills, MCP servers, and agent plugins are the three ways to extend an AI agent in 2026. A skill is a folder of instructions the agent reads when a task matches. An MCP server is a program the agent calls to reach a tool or a dataset. A plugin is the package that installs a skill and its MCP servers together, in any client that speaks the format.
People still search "Claude Skills vs MCP" because Anthropic shipped Skills in October 2025 and the two got filed as rival products. The skill is the procedure. MCP is the connection.
What is an Agent Skill?
An Agent Skill is a recipe card. The directory holds a SKILL.md file: YAML frontmatter (name and description required), then Markdown instructions. Optional scripts/, references/, and assets/ folders hold code, extra docs, and templates. The spec caps name at 64 characters and description at 1,024 characters, and the name has to match the folder.
Anthropic introduced this on October 16, 2025 for when the model can already act and the missing piece is your procedure. Their PDF skill is the pattern: forms.md stays closed until the agent is filling a form, and a script extracts the fields so the PDF never has to sit in the window. A skill can run code, which means it can also be written to send data out. Read the scripts before you install one you didn't write.
What is an MCP server?
The Model Context Protocol is how an AI app pulls context from another program. The server is a phone jack. The host (Claude Code, VS Code, anything else that speaks MCP) opens one client per server. A server can expose three things:
Tools, functions the agent can run: query a database, file a ticket, search Sentry.
Resources, data it can read: a schema, a file, a record.
Prompts, reusable templates for talking to those tools.
Local servers use stdio. Remote servers use Streamable HTTP. The protocol carries the context. The host decides what the model sees.
MCP launched in November 2024. By the 2026-07-28 release, Anthropic said the SDKs were past 400 million downloads a month, up 4x on the year, and Claude's connector directory listed over 950 servers.
What is an agent plugin?
An agent plugin is the shipping carton: a directory with a plugin.json manifest ($schema and name are the minimum), a skills/ folder of Agent Skills, and an optional mcp.json that names the MCP servers to connect. Version 1.0.0, published August 6, 2026, allows those two component types and leaves commands, hooks, and agents to each client.
Vercel started it. AWS, Anysphere (Cursor), GitHub, Microsoft, OpenAI, and Vercel wrote the spec. The steering committee is AWS, Cursor, Microsoft, OpenAI, and Vercel. At launch it loaded in ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. FOMA's August 13, 2026 issue covered that week: the labs arguing over whose model was smarter had agreed on an install format.
How does each one load, and what does it cost in context?
A skill stays small until the task matches. The spec calls it progressive disclosure: at startup the agent loads only each skill's name and description, about 100 tokens of metadata. When the description fits, it reads the full SKILL.md. The spec recommends keeping that body under 5,000 tokens. Scripts can run outside the window.
An MCP server costs whatever tool list the host pastes in. The 2026 client guide says a naive host dumps every connected tool schema into the prompt at the start. A few tools is fine. Dozens of servers can take most of the window before anyone types. Anthropic's November 2025 example: GitHub, Slack, Sentry, Grafana, and Splunk, 58 tools, about 55,000 tokens before the conversation. The same month, loading tools as files the agent opens on demand cut one run from 150,000 tokens to 2,000 (98.7%). Those figures are from late 2025. The 2026 guide says to switch to search-then-load around 1% to 5% of the window, and to connect a server when the skill that names it runs. Intermediate results come back into context unless the host runs the chain in code.
A plugin costs whatever is inside it: skill metadata until one triggers, then that host's tool-list bill for any servers it connects.
MCP's Skills extension (io.modelcontextprotocol/skills, on the 2026-07-28 spec) can serve those same folders as skill:// resources. Same file format, over the protocol, for when the manual should ship with the server.
Are Agent Skills still Claude-only?
No. They are an open standard. Anthropic published the format for other products on December 18, 2025. It lives at agentskills.io, and the client list there includes Claude and Claude Code, ChatGPT and Codex, Cursor, VS Code, GitHub Copilot, Gemini CLI, OpenCode, and Goose.
Claude's marketplace calls both open standards Anthropic pioneered: more than 2,000 connectors and plugins, including Atlassian and Salesforce, plus a shelf where committed Anthropic spend can buy agents from CrowdStrike, Cursor, and others. FOMA covered that catalog on October 1, 2026. People still type "Claude Skills." The format is shared.
What do people mean by Claude tools vs MCP?
Three layers, one search query. In Claude, a tool is anything the model can call: a built-in (bash or file edit), a tool you declared on the API, or a tool that arrived from an MCP server. MCP is one supply route, from a process on your machine or on someone else's. A skill is the instructions for which tools to call, and in what order. Anthropic's original Skills post said the two were built to be used together.
What changed in MCP in 2026?
MCP 2026-07-28 is the largest revision since launch. FOMA's July 30 issue called it Anthropic rebuilding MCP and turning Claude Code into a platform.
The core is stateless, per the architecture docs. The initialize handshake and the Mcp-Session-Id header are gone. Every request carries version and capabilities in _meta, so any instance can answer it, including on serverless and edge. Want state across calls? The server mints an id, and the model passes that id back as a normal argument.
Tasks and MCP Apps are official extensions. A task is a handle from tools/call that the client polls (tasks/get), feeds (tasks/update), or stops (tasks/cancel). An MCP App is HTML the host renders in a sandboxed iframe, with clicks on the same audit path as a direct tool call.
Auth matches production OAuth 2.0 and OpenID Connect, including Entra and Okta. Roots, sampling, and logging are deprecated (pass paths as arguments, call a model API yourself, log to stderr or OpenTelemetry), with at least twelve months before removal, and the old methods still work in this version. On Claude: Apps in the chat, org-wide connectors through the company identity provider, and MCP tunnels in research preview so a private server needs no public URL.
When should I use a skill, an MCP server, or a plugin?
Use a skill for the house procedure: a review checklist, a migration recipe, a changelog format, or a document job where a script does the mechanical part (the PDF skill).
Use an MCP server when the filesystem cannot see the system. A database, with the schema as a resource and the query as a tool. A live API in the shape the docs keep using: Salesforce, Sentry, a Figma canvas, Zoom notes in the chat. A long job that should return a task handle. A server on a private network, which is what the tunnel preview is for.
Use a plugin when someone else has to install both: a refunds skill plus the billing server, one plugin.json, in Codex, Cursor, Copilot, Kiro, and VS Code. Solo, with the skill already in the repo, skip the manifest.
Which one should I build?
If the gap is "the model doesn't know how we do this," write a SKILL.md and commit the folder. A checklist does not need a server.
If the gap is "the model can't reach the system," write an MCP server, with a few tight tools rather than a 58-tool drawer.
Name those servers in the skill that needs them, and connect them when the skill runs.
Someone else has to install the skill and the server? Wrap them in an Agent Plugin.
Only enabling someone else's connector? Read what it can touch. The marketplace lists 2,000+.
Verdict
You will own a skill and a server once the agent has to follow a house procedure and touch a system it cannot see. The skill is the cheap one: its body stays out of the prompt until the task matches. The server can quietly spend tens of thousands of tokens on tool schemas if the host loads them all up front. The plugin is how a second person installs that pair. FOMA AI News is the weekly filter for everything else on the stack.