> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zeroruntime.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Setup with MCP and Skills

> Connect the Zero Runtime MCP server and Agent Skill to your AI coding agent so it builds with the current docs instead of guessing.

Two pieces make an AI coding agent good at Zero Runtime:

* **The MCP server.** A live search over these docs, hosted at `https://mcp.zeroruntime.ai`. The agent looks up real API details mid-task instead of recalling them.
* **The Agent Skill.** The procedure for building on Zero Runtime: scoping the agent, wiring auth, and picking a pipeline.

Install both. The skill knows *what* to build; the MCP server tells it the *exact* method name.

## Install the MCP server

Pick your client. Most clients need a restart before a new server loads.

<Tabs>
  <Tab title="Claude Code">
    Run this from your project, or add `--scope user` to make it available everywhere:

    ```bash theme={null}
    claude mcp add --transport http zeroruntime https://mcp.zeroruntime.ai
    ```

    Confirm it registered with `claude mcp list`.
  </Tab>

  <Tab title="Cursor">
    Add the server to `~/.cursor/mcp.json` for every project, or `.cursor/mcp.json` for this one:

    ```json theme={null}
    {
      "mcpServers": {
        "zeroruntime": {
          "url": "https://mcp.zeroruntime.ai"
        }
      }
    }
    ```
  </Tab>

  <Tab title="VS Code">
    Create `.vscode/mcp.json` in your workspace. Copilot picks it up once the file is saved.

    ```json theme={null}
    {
      "servers": {
        "zeroruntime": {
          "type": "http",
          "url": "https://mcp.zeroruntime.ai"
        }
      }
    }
    ```
  </Tab>

  <Tab title="Codex">
    Add the server to `~/.codex/config.toml`:

    ```toml theme={null}
    [mcp_servers.zeroruntime]
    url = "https://mcp.zeroruntime.ai"
    ```
  </Tab>

  <Tab title="Windsurf">
    Add the server to `~/.codeium/windsurf/mcp_config.json`:

    ```json theme={null}
    {
      "mcpServers": {
        "zeroruntime": {
          "serverUrl": "https://mcp.zeroruntime.ai"
        }
      }
    }
    ```
  </Tab>

  <Tab title="Antigravity">
    Add the server to `~/.gemini/config/mcp_config.json` for every project, or `.agents/mcp_config.json` for this one:

    ```json theme={null}
    {
      "mcpServers": {
        "zeroruntime": {
          "serverUrl": "https://mcp.zeroruntime.ai"
        }
      }
    }
    ```

    <Note>
      A remote server must use `serverUrl`. The legacy `url` and `httpUrl` fields are not supported. Type `/mcp` in the prompt panel to manage servers from inside the CLI.
    </Note>
  </Tab>

  <Tab title="Cline">
    Open the MCP Servers panel, choose to edit the configuration, and add the server to `cline_mcp_settings.json`:

    ```json theme={null}
    {
      "mcpServers": {
        "zeroruntime": {
          "type": "streamableHttp",
          "url": "https://mcp.zeroruntime.ai"
        }
      }
    }
    ```
  </Tab>

  <Tab title="Zed">
    Add the server under `context_servers` in your Zed `settings.json`:

    ```json theme={null}
    {
      "context_servers": {
        "zeroruntime": {
          "source": "custom",
          "url": "https://mcp.zeroruntime.ai"
        }
      }
    }
    ```
  </Tab>
</Tabs>

### Verify the connection

Ask your agent:

> Search the Zero Runtime docs for how to generate an auth token, and tell me which page it came from.

A connected agent replies with a real `docs.zeroruntime.ai` URL it just fetched. If it answers from memory, cites no page, or says it has no such tool, the server has not loaded. Restart the client and recheck the file you edited.

## Install the Agent Skill

One skill, `zeroruntime`, covering voice agents, telephony, and the auth each needs.

```bash theme={null}
npx skills add ZeroRuntimeAI/skills
```

`npx skills` is agent-agnostic. It asks which agents to install into. Pick the ones you use, then restart them. To place it by hand instead, copy the `zeroruntime` folder into your agent's skills directory, for example `~/.claude/skills/zeroruntime`.

<Note>
  The skill covers voice agents today. For Physical AI the MCP server still
  searches the robotics pages, so the agent works from the real docs. It just
  has no pre-written procedure to follow, which makes a specific prompt matter
  more.
</Note>

## Sample prompts

Once the server is connected, describe what you want in plain language. The
agent searches these docs as it works, so it uses the real class and method
names instead of guessing them.

### Build a voice agent

> Using the Zero Runtime docs, build me a Python voice agent that answers our
> support line. Use Deepgram for speech to text, Gemini for the model, and
> Cartesia for the voice. Give it a tool that looks up an order by number, and
> have it hand off to a human when the caller asks for one. Show me how to run
> it locally first.

> Add telephony to the agent I just built so it answers a real phone number.
> Walk me through what I need to set up on the provider side.

### Build a physical AI teleoperation setup

> Using the Zero Runtime docs, set up remote teleoperation for my hardware. The
> follower is an SO-101 arm on a Raspberry Pi with one USB camera, and I want to
> drive it from my laptop. Write both scripts, set safe joint limits, and record
> the session so I can export a LeRobot dataset later.

> My robot is not an SO-101. Show me how to write a custom follower adapter for
> it, and which methods I have to implement.

<Tip>
  Name the hardware, the providers and the language you want. The more specific
  the prompt, the less the agent has to assume.
</Tip>

## Skill or MCP server

They do different jobs and work best together.

| | Agent Skill | MCP Server |
| - | - | - |
| What it is | The procedure the agent follows | The lookup the agent calls |
| What it holds | How to scope the build, wire auth, pick plugins | The current docs, searchable |
| When it runs | At the start of a Zero Runtime task | Whenever the agent needs an API detail |

## Troubleshooting

<AccordionGroup>
  <Accordion title="The agent says the tools are unavailable">
    Restart the client. Several of them only initialise MCP servers at startup.
  </Accordion>

  <Accordion title="A search comes back empty">
    Search for the full name as it appears in the docs, for example a plugin's provider name, rather than an abbreviation.
  </Accordion>

  <Accordion title="A result is wrong or out of date">
    Ask the agent to send feedback through the server, including the page and the query you used. That reaches the team maintaining the index.
  </Accordion>
</AccordionGroup>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.