Agent skills vs MCP: what each does and when to use both

Agent skills and MCP servers solve different problems. An agent skill is packaged know-how: a SKILL.md file of instructions, plus optional scripts and reference files, that the agent reads when a request matches the skill's description. An MCP server exposes tools, named actions with an input schema that the agent calls to act on another system. The skill tells the agent how to do a job; the tools do parts of it, and one agent can use both.

Skill behavior below comes from Anthropic's Agent Skills overview and tool behavior from the MCP specification's Tools page, both read on 2026-09-27. Sume's examples come from its Format API and MCP overview docs.

What is an agent skill?

In Anthropic's words, each skill packages instructions, metadata, and optional resources such as scripts and templates, which Claude uses automatically when relevant. A skill is a folder; in Claude Code, personal skills live in ~/.claude/skills/ and project skills in .claude/skills/. Its SKILL.md starts with YAML frontmatter whose required fields are name and description, and the description has to say both what the skill does and when to use it.

Skills load in stages, which Anthropic calls progressive disclosure:

• Metadata, always: name and description load at startup, about 100 tokens per skill.
• Instructions, when triggered: Claude reads the SKILL.md body, under 5k tokens, only when a request matches the description.
• Resources, as needed: other files load only when read. Scripts run through bash, and only their output enters context.

What is an MCP tool?

The Model Context Protocol lets servers expose tools that language models can invoke to interact with external systems, such as querying databases, calling APIs, or performing computations. Each tool has a unique name, a description, and an inputSchema, a JSON Schema for its parameters. A client discovers tools with a tools/list request and runs one with tools/call.

The spec calls tools model-controlled: the model can discover and invoke them automatically. For trust and safety, it says there SHOULD always be a human in the loop with the ability to deny tool invocations, and that applications should show which tools are exposed and ask for confirmation.

How do skills and MCP tools differ?

When should I use a skill, and when an MCP server?

• Use a skill for know-how: a procedure, a house style, or a checklist the agent should apply the same way every time. Skills load on demand, so you don't repeat the same guidance in every conversation.
• Use an MCP server for actions on another system: a database, an API that needs its own credentials, a paid job.
• Use both when a task needs judgment and actions. The skill says which tools to call, in what order, and what to check; the tools make the calls.
• Check where the skill runs. On the Claude API, Anthropic says skills have no network access and cannot make external API calls; in Claude Code, they have the same network access as any other program on your computer.

How does Sume use skills and MCP together?

Sume splits the work the same way. A Sume Format is a saved recipe: a SKILL.md body plus reference files that carry the how, such as house style and quality bar, while each call's instruction and input carry the what. In a run, the agent follows the recipe and calls generation tools, which attach to each turn as Sume MCP tools. Sume's docs give advice that matches progressive disclosure: keep SKILL.md a short index and push detail into references/*, which cost nothing until opened. A Format uses the SKILL.md convention, but it runs in a fresh Sume sandbox, not inside your own agent; How to write a SKILL.md file for a Sume Format covers writing one.

For your own agent, the tools side is Sume's hosted MCP server at https://mcp.sume.com/mcp, which lets agents call Sume tools such as generate_video and jobs_wait without wrapping the HTTP API. Its tools wrap selected API capabilities, not all of them. The Sume CLI also ships a packaged skill: sume skills install writes it into .agents/skills or .claude/skills; see Claude skill to generate video. Sume's basics page says the CLI and hosted MCP still work but are not the primary path today; backends call Formats over HTTP, as MCP vs CLI vs API for AI agents explains.