Skills and MCP

Skills and MCP servers are the two main ways to teach an agent new tricks once it is running. A skill gives the agent a written routine for one kind of job, while an MCP server gives it new tools, such as a link to another app or service. This page explains each one, shows which of the ten agents in this guide support them, and walks through adding one of each to Agent Zero. Everything here was checked on 14 November 2026.

Both are code or instructions written by other people, so read the safety section before you add anything you did not write yourself.

What a skill is

A skill is usually a folder with a file named SKILL.md inside it. The top of that file holds a short name and a one-line description, and the rest is plain instructions, written the way you would brief a new colleague. Some skills also carry scripts, templates, or reference files in the same folder.

The agent does not read every skill all the time. At first it sees only each skill's name and description, and it reads the full instructions when your request matches one. That lets it keep many skills on hand without filling the prompt, which is also why a clear description matters so much.

One format, several agents. Agent Skills is an open format for these folders, which Anthropic first developed and then released as a standard. Its site, agentskills.io, lists Goose, Hermes Agent, and OpenClaw among the agents that support it. Agent Zero's own docs say its SKILL.md format works with Goose, Claude Code, Cursor, and others, and that skills from those can be copied straight into its skills folder. Atomic Agent, NanoClaw, QwenPaw, Eigent, and Everfern also use SKILL.md files, but this guide could not confirm that they follow the whole standard, so a skill written for one agent may need small changes in another.

What MCP is

MCP, short for Model Context Protocol, is a standard way to plug outside tools into an agent. The program that offers the tools is called an MCP server, and an agent connects to every server the same way, whoever wrote it. One server might tell the time, another might read a database, and a third might search a website, and their tools then sit beside the agent's built-in ones.

MCP servers come in two kinds, and the difference matters for safety.

• A local server runs on your own computer. The agent starts it with a command, such as uvx or npx followed by a package name, and talks to it directly. The MCP docs say the agent and a local server "run side by side with the same privileges," so the server can reach whatever the agent's account can.
• A remote server runs on someone else's computer, and the agent reaches it at a web address. It runs no code on your machine, but it sees whatever the agent sends it, and it often needs a key or a sign-in to your account with that company.

Where "your account" ends depends on the agent. Agent Zero and NanoClaw start local servers inside their containers, so such a server reaches the container rather than your real files. Most other agents in this guide run on your real computer, so a local server they start runs as you.

Skills or MCP: which to use

A skill tells the agent how to do a job with what it already has, and an MCP server gives it something new to do the job with. Agent Zero's docs sum up the split as "How to do X" for a skill and "Do X now" for a tool.

Pick a skill when you keep pasting the same instructions, such as "summaries in five plain bullets" or the checklist for a weekly report. Pick an MCP server when the job needs something the agent cannot reach today, such as a service that publishes its own MCP server. If the agent already has a tool that does the job, use that first; Agent Zero's MCP guide gives the same advice about its built-in browser. Many jobs need neither, because a clear request does the work, as Prompting strong and weak models explains.

Which agents support them

Each agent's name links to the page in this guide that covers its skills and MCP setup. "Not documented" means this guide could not confirm the feature in that agent's official docs or code.

A few differences matter more than the table shows.

• Hermes Agent writes and rewrites its own skills by default, unless skills.write_approval: true is set in config.yaml.
• Atomic Agent, Hermes Agent, OpenClaw, and QwenPaw scan skills from their catalogs, but none of their docs treat a clean scan as proof of safety. QwenPaw's scanner also starts on Warn, which lets a risky skill through with only a warning.
• Goose checks local extensions for known malware, but its docs say the check covers only npx and uvx packages, and that extensions still install when the check service is down.
• Everfern's MCP tool is something the agent calls itself, and this guide could not confirm whether Everfern asks you before it connects a server.

How to add a skill

These steps work for any agent in the table.

1. Decide what the skill is for. One skill should cover one kind of job, such as a summary format or a weekly checklist.
2. Write it yourself, or get it from a trusted place. That means the agent's official project or a well-known publisher, and a catalog listing on its own does not make a publisher trusted.
3. Read every file before you add it. Read the whole SKILL.md and any scripts beside it. Reject it if it names folders, accounts, or websites the job does not need, or tells the agent to keep anything from you.
4. Put it where your agent looks. Use the folder, screen, or command from the table above.
5. Check that the agent sees it, in its skill list, selector, or list command.
6. Try it once on a harmless job while you watch, using a copy of a file rather than the only one you have.

Worked example: a plain-summary skill in Agent Zero

This skill is only instructions, with no scripts, so it cannot do anything the agent could not already do. It asks for one short summary file and tells the agent to install nothing and stay off the web.

1. On your computer, open the workspace folder from Step 2, which is ~/agent-zero/<instance>/usr. If there is no folder named skills inside it, make one.
2. Inside skills, make a folder named plain-summary. The folder name must match the name line in the file below.
3. Open Notepad on Windows or TextEdit on a Mac, paste the text below, and save it in plain-summary as plain text named SKILL.md. Check that the saved name does not end in .txt.

---
name: plain-summary
description: Write a short plain-language summary of one text file in the workspace. Use when the user asks for a plain summary of a file.
---

# Plain summary

When the user asks for a plain summary of a file:

1. Read only the file the user names, inside the workspace.
2. Write at most five short bullet points in plain words, with the most useful fact first.
3. Save them in the same folder as a new file whose name ends in -summary.txt.

Do not install anything, open websites, sign in to anything, or change or delete any other file.

4. In Agent Zero, open a new chat, click the + button in the chat input, and click Skills. Search for plain-summary and click it to pin it. If it is not in the list, click Restart in the sidebar and look again, because Agent Zero's docs say skill changes take effect after a restart.
5. Give it a first job. This one uses shopping-notes.txt from Step 2, but you can name any text file you have copied into the workspace.

Use the plain-summary skill on shopping-notes.txt in the workspace. Follow the safety rules: stay in the workspace and do not sign in or buy anything.

What you should see a new shopping-notes-summary.txt beside shopping-notes.txt, holding a few short bullet points, with no other file changed.

6. When you no longer want it pinned, open Skills again and use the remove button under Active skills. Agent Zero can still load it by itself when a request matches. Its docs advise keeping the pinned list short, because "A lighter prompt is usually easier for the agent to follow".

The skill lives in the workspace, which the agent can also write to. If it ever starts behaving differently, open SKILL.md again and check that the text is still yours. To remove it completely, delete the plain-summary folder.

How to add one MCP server

These steps work for any agent that supports MCP.

1. Pick one server for one job. Start with a server from the official MCP servers collection on GitHub, or one published by the company whose service it connects to.
2. Read its page first. Note which tools it offers, whether any of them write, send, or delete, whether it needs a key, and whether it runs on your computer or at a web address.
3. Paste its settings where your agent takes them. Use the place from the table above. Agent Zero, QwenPaw, and OpenManus all take the same JSON shape: a block named mcpServers with one entry for each server.
4. Check that it connected. Most agents show a status and a count of the server's tools.
5. Turn off the tools you do not need, where your agent allows it, and keep approvals on.
6. Test it with a harmless question while you watch, and read which tool the agent called.
7. Remove it when you stop using it.

Worked example: the time server in Agent Zero

The time server comes from the official modelcontextprotocol/servers collection on GitHub. Its tools, such as get_current_time and convert_time, only read the clock and do time-zone math. It touches no files, needs no key, and signs in to nothing, which makes it a sensible first server even though it is not very useful on its own.

1. In Agent Zero, click Settings in the sidebar, open the MCP/A2A tab, find External MCP Servers, and click Open.
2. Paste the block below. It is the time server's own example, and it has the same shape as the command-based example in Agent Zero's MCP guide. If the editor already lists servers you want to keep, add only the "time" entry inside the existing "mcpServers" braces, with a comma after the entry before it.

{
  "mcpServers": {
    "time": {
      "command": "uvx",
      "args": ["mcp-server-time"]
    }
  }
}

3. Click Apply now. The first start can take a little longer, because Agent Zero's guide says a uvx server may need to download first.
4. Look below the editor for the server's status.

What you should see a server named time, a count of its tools, and a green status.

5. Ask a question that needs it. Name a city, because the server runs inside Agent Zero's container, whose clock may be set to a different time zone from yours.

Use the time tool to tell me the current time in Tokyo and in London. Do not open any website or change any file.

What you should see the chat shows a call to a time tool, and the answer matches a world clock.

6. If the agent says it cannot use the tool, the agent profile may be blocking it. Open the profile with Edit agent, go to MCPs, and set the time tools to On.

To remove the server later, open the same editor, delete the "time" entry, and click Apply now. Once you have seen how a server connects, the same steps work for one that does a real job.

Safety: code and instructions from strangers

Every skill and MCP server you did not write gives its author a say in what your agent does. Install only from the agent's official project or a well-known publisher, read a skill before you add it, and start with tools that only read.

A skill steers the agent, and its scripts run as the agent. Its instructions shape every job that loads it, and a skill can load by itself whenever a request seems to match. Atomic Agent's docs say skill scripts run with the agent's full permissions and are not sandboxed, and OpenClaw's docs say a bad skill can read your keys. Catalog scans help, but Atomic Agent's docs admit a scan "cannot catch a script that is merely well-written and hostile".

A local MCP server runs with your account's rights. On an agent that runs on your real computer, a local server can read, change, or send anything your account can, and the approval prompt does not wall it off. Many setup examples start a server with npx -y, and the -y tells npm to download and run the package without asking first. Agent Zero and NanoClaw narrow that reach to their containers, which is one more reason to keep Host access off.

Tool descriptions are a path for prompt injection. When an agent connects to a server, the server sends a name and description for each tool, and the model reads them as instructions. The MCP docs call this "an injection surface that install-time review does not cover". They add that one bad server "can influence how the agent uses the tools of every other server". Read the tool list once after you connect a server, and remove it if any description asks for things unrelated to its job. Results a tool brings back can carry a prompt injection too, so a server that reads email or web pages is riskier than one that tells the time.

Too many tools cost tokens and confuse weaker models. Every connected tool adds its name and description to what the model reads, so each server you add makes every request bigger and more expensive. Goose's docs say it works best with fewer than 25 tools turned on, and Agent Zero's skills guide says to keep the pinned list short. A small or local model also has a harder time choosing well from a long list, so with a weak model, turn on only the tools today's job needs.

Start read-only. Begin with a server like the time server, whose tools only read. Add one that writes, sends, or deletes only when a job needs it. Then keep its approval on, and connect it to an account made for the agent rather than your main one.

Trade-offs: safety versus capability

Installing a community skill. A skill someone else wrote, from a catalog such as ClawHub or the QwenPaw market, or from a page on GitHub.

• You gain: a tested routine for a kind of job, such as slides or a report format, without writing it yourself.
• You risk: its instructions steer every job that loads it, and any scripts run with the agent's permissions. A catalog scan can miss a hostile script that is well written.
• Makes sense when: it comes from the agent's official set or a publisher you recognize, you have read every file in it, and it only reads and writes in the workspace. A skill that mentions your email, passwords, or folders outside the workspace does not make sense for a beginner.

Adding a local MCP server. A server the agent starts on your computer with a command such as uvx or npx.

• You gain: a new kind of tool with no account and no company in between, such as the time server, or one that reads a single folder you choose.
• You risk: it is a program running with your account's rights, or the container's on Agent Zero and NanoClaw, and its tool descriptions go straight to the model.
• Makes sense when: it comes from the official servers collection or the service's own maker, its tools only read, and you watch the first job that uses it. Turn it off when the job is done.

Adding an MCP server that can write or send. A server whose tools can change files, post, send email or messages, or act in one of your accounts.

• You gain: the agent can finish a job instead of drafting it, such as filing a note in an app you use or sending a reply.
• You risk: a prompt injection in any page or message the agent reads could use those tools, and a sent message or a deleted record usually cannot be undone.
• Makes sense when: the account is one made for the agent, approval stays on for every write or send, and you are watching. Connecting your main email, bank, or shopping account does not make sense for a beginner, and neither does running such a server on a schedule.

The three questions behind these choices are on Safety versus capability.

Sources

Checked 14 November 2026.

Shared standards:

• Agent Skills overview, for skills as folders with a SKILL.md file, loading on demand, and the format's origin at Anthropic.
• Agent Skills specification, for the required name and description and the rule that the name matches the folder.
• Agent Skills client showcase, for Goose, Hermes Agent, and OpenClaw.
• MCP architecture, for MCP servers and local versus remote servers. Also checked through Context7.
• MCP local server security, for "same privileges," tool descriptions as an injection surface, and one server influencing the others.
• MCP time server README, for its tools and the uvx example.
• npm, npx, for -y skipping the prompt before a package is downloaded and run.

Agent Zero:

• Skills guide, for +, Skills, Active skills, and keeping the pinned list short.
• Contributing Skills, for usr/skills, the cross-platform claim, "How to do X" and "Do X now", and changes taking effect on restart.
• Skill loader code, for usr/skills being one of the folders Agent Zero searches.
• Usage guide, for Restart in the sidebar.
• MCP Setup guide, for MCP/A2A, External MCP Servers, Apply now, the status check, Edit agent and MCPs, uvx downloads, and servers that run inside the container.
• Python install script, for uv being installed in Agent Zero's container.
• Agent Zero research notes.

Other agents:

• Atomic Agent: Skills and MCP client.
• Hermes Agent: Skills and MCP, plus the Hermes research notes.
• OpenClaw: Creating skills, Skill security audits, and MCP.
• Goose: Agent skills, Using extensions, Known issues, and Tool permissions, for fewer than 25 tools.
• NanoClaw: Skills overview, for host skills and container skills, and Give agents tools.
• QwenPaw: Skills, MCP, and Security, for the Skill Scanner.
• Eigent: Agent skills and Connectors.
• Everfern: Settings screen source and MCP code.
• OpenManus: README, for the Browser Use skill, and settings loader, for config/mcp.json.
• Research notes on the other agents.

Next

The next page is Persona and rules.