Skills, MCP and CLI: What Is the Difference for an Agent?
Display is our product. We checked the Display facts against the CLI reference and MCP tools on 14 October 2026.
Display is our product. We checked the Display facts against the CLI reference and MCP tools on 14 October 2026.
The short answer
| Skill | MCP server | CLI | |
|---|---|---|---|
| What it is | A folder with a SKILL.md file of instructions | A server that gives named tools | A program that the agent runs in a shell |
| What the agent does | Reads the instructions, then uses tools or commands | Calls a tool with structured input | Runs a command and reads the output |
| Needs a shell | Depends on the skill | No | Yes |
| Works in chat apps | Some | Yes, through connectors | No |
| Display example | The Display skill, added to the agent | The Display MCP server at https://api.display.dev/v1/mcp | The Display CLI |
A skill does no work by itself. In Claude Code, a skill is a directory with a SKILL.md file. The host keeps the description of each skill in context. It loads the full text only when it uses the skill (Claude Code docs, checked 14 October 2026). The skill then tells the agent which tool or command to use, and in which order.
The same task with each one
The task: publish a report for the company, then update it after a comment.
With the CLI
The agent runs the Display CLI in its shell. It publishes the file for the company and gets the URL. Later it lists the comments and reads the artifact metadata. Then it publishes the next version at the same URL, based on the version that it edited. Read CLI commands for the commands.
With MCP
The agent calls publish with the content and the visibility. For an update, it sends base_version, the version that it edited. Later it calls list_comments, then edit for one exact change, then resolve_thread. The authenticated server has 31 tools. Remote MCP keeps source reads small: search finds text and read returns that range. Read MCP tools.
With the skill
You ask: "Publish this report for the company. Fix the open comments." The skill tells the agent to publish, read the comments, make an exact edit, reply and resolve. The skill uses MCP or the CLI to do the work. Read Agent skill reference.
Which one to choose
- Claude on the web, desktop or mobile. Use the connector from the Claude directory. It is MCP with OAuth.
- Cursor, Codex, Claude Code. Add the skill. It sets up the connection and teaches the workflow.
- A script or a CI job. Use the CLI with an API key.
- No account yet. Send
curl -X POST https://api.display.dev/v1/public/artifacts -F file=@report.html. You get a preview URL for 30 days and a claim URL. Read Publish without an account.
When another approach is better
If you have only a chat window and cannot add a connector, the CLI will not run. Upload the file in the dashboard. If the agent has no shell and no MCP, use web upload.
FAQ
Is a skill the same as an MCP server?
No. A skill is a set of instructions. An MCP server runs tools. A skill can tell the agent to call an MCP server.
Do I need the skill if I have the MCP connection?
No. The tools work without the skill. The skill adds the workflow, for example how to update the right version and how to reply to comments.
Which one is the fastest to start?
A publish without an account needs no sign-in. Ask the agent to publish the file without an account. It gives you a preview URL and a claim URL.
Do all three publish to the same URL?
Yes. An update from any of the three creates the next version at the same URL.
Free tier. No credit card. One-time password auth for viewers on free, Google + Microsoft SSO on Pro (€49/month flat).