MCP Servers for cloud agents

Cloud agents can call external tools through Model Context Protocol (MCP) servers. This lets agents reach beyond the terminal to automatically interact with systems like GitHub, dbt, Sentry, or any custom internal service, whenever the workflow requires it.

When to use MCP servers

Add MCP servers to a cloud agent when it needs to:

• Read from or write to an external API (issue trackers, monitoring tools, cloud services)
• Call local processes that expose MCP endpoints
• Use internal developer tools that you’ve wrapped in an MCP interface

The agent calls MCP tools automatically based on what the task requires, without the need for explicit instruction.

How MCP configuration works

You can supply MCP configuration in two ways:

• At run time — pass --mcp when calling oz agent run or oz agent run-cloud. See MCP Servers in the CLI reference for the full syntax.
• In an agent config file — define mcp_servers directly in a YAML or JSON agent config file (passed with -f / --file). This is the recommended approach for repeatable workflows.

Configuration schema

Each MCP server entry is keyed by a name you choose. A server config must have exactly one transport type:

Transport	Field(s)	When to use
Warp-shared server	warp_id	Reference an MCP server already configured in Warp by its UUID
Stdio (local process)	command, args	Launch a local executable as an MCP server
Streamable HTTP / SSE	url	Connect to a remote or locally hosted MCP endpoint

Supported fields

• warp_id — UUID of a Warp-shared MCP server (find UUIDs with oz mcp list or from Settings > Agents > MCP servers)
• command — Executable to launch (stdio transport)
• args — Arguments passed to command (only valid with command)
• env — Environment variables passed to the process (only valid with command)
• url — HTTP or HTTPS endpoint URL (streamable HTTP or SSE transport)
• headers — HTTP headers sent with requests (only valid with url)

You may define any number of MCP servers in a single config.

Example configuration

{
  "github": {
    "url": "https://mcp.example.com/github"
  },
  "dbt": {
    "command": "uvx",
    "args": ["dbt-mcp"],
    "env": {
      "DBT_HOST": "https://example.us1.dbt.com",
      "DBT_SERVICE_TOKEN": "${DBT_SERVICE_TOKEN}"
    }
  }
}

Using MCP servers in an agent config file

For repeatable cloud agent workflows, declare your MCP servers inside the agent config file passed to -f / --file:

{
  "name": "my-production-agent",
  "model_id": "claude-sonnet-4",
  "system_prompt": "You are a helpful assistant focused on backend development.",
  "environment_id": "SVhg783GBFQHk1OfdPfFU9",
  "mcp_servers": {
    "github": {
      "url": "https://mcp.example.com/github"
    },
    "dbt": {
      "command": "uvx",
      "args": ["dbt-mcp"],
      "env": {
        "DBT_HOST": "https://example.us1.dbt.com",
        "DBT_SERVICE_TOKEN": "${DBT_SERVICE_TOKEN}"
      }
    }
  }
}

Pass this file when running a cloud agent:

oz agent run-cloud --environment <ENV_ID> -f my-agent-config.json --prompt "Check for regressions in the last deploy"

Requirements and defaults

• MCP configuration must be valid JSON, or YAML when embedded in a broader agent config file.
• If mcp_servers is omitted, the agent runs with no MCP servers enabled.
• Each server name must be unique and non-empty.
• The warp_id transport is validated against your Warp account. Referenced servers must be accessible to you.

Limitations

Caution

Warp does not currently support OAuth-based MCP servers for cloud agents. This means MCP servers that require browser-based authentication, like some hosted Figma configurations, cannot be used directly.

As a workaround, you can pass Figma mockups as image context to the agent, which can then build and test UI against those images.

Learn more

• MCP Servers (CLI reference) — how to pass MCP configuration using the --mcp flag
• Model Context Protocol (MCP) — configuring MCP servers in Warp for local agents
• Environments — set up the runtime context (repo, image, startup commands) for cloud agent tasks
• Secrets — store and inject credentials into agent runs safely