Update MCP

Manage MCP Server Templates

MCP (Model Context Protocol) server configurations define external tool servers available to AI tools. They use JSON format (not markdown), similar to hooks.

Finding Existing MCP Configs

List files in <%= mcpTemplatePath() %>/ to discover existing MCP templates (.json files). Read them to understand what servers are already configured.

Note: Servers from multiple files are merged by server name during generation, and the first declaration of a name wins (templatesDir before additionalTemplateDirs, in order). You can organize servers by concern (e.g. github.json, databases.json).

Deciding What to Do

• Create new file: for an entirely new server or group of related servers
• Update existing file: modify server config, add servers to an existing file
• Merge strategy: since servers merge by name, you can split them across files for better organization
• Delete: remove an MCP config file when the server is no longer needed

Creating a New MCP Config

1. Create a .json file in <%= mcpTemplatePath() %>/ with a descriptive name (e.g. github.json)
2. Use the standard MCP JSON structure

JSON Structure

{
  "mcpServers": {
    "server-name": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@some/mcp-server"],
      "env": {
        "API_KEY": "${API_KEY}"
      }
    }
  }
}

Server Fields

See the MCP Servers → Server Fields section in <%= instructionPath('uac-template-guide') %> for the complete reference, including all 13 Codex-specific fields (cwd, envVars, enabledTools, disabledTools, bearerTokenEnvVar, envHttpHeaders, startupTimeoutSec, startupTimeoutMs, toolTimeoutSec, enabled, required, oauthResource, scopes, experimentalEnvironment).

Key cross-target fields:

A server must have either command (stdio) or url (HTTP). If neither is present after per-target override resolution, the server is dropped for that target.

Codex notes

• Codex output emits TOML (.codex/config.toml) merged into the [mcp_servers.*] table. Other top-level keys in config.toml ([profiles.*], [model_providers.*], [permissions.*], personality, etc.) are preserved across regenerates — uac only owns [mcp_servers.*].
• Auth model differs: Codex uses bearerTokenEnvVar + oauthResource + scopes instead of Claude's oauth object or Cursor's auth object. Claude/Cursor auth fields are dropped with a warning when emitting for Codex; use per-target overrides for Codex-specific auth.
• Per-server tool restriction lives at the MCP server level via enabledTools (allowlist) and disabledTools (denylist) — agent-level tools fields are not applicable for Codex.

Variable Interpolation

Use {{variableName}} syntax to reference variables from the config file. Variables support typed resolution: when the entire JSON value is "{{varName}}", it resolves to the raw typed value (array, object, number, boolean). When embedded in other text, it does string interpolation.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": "{{playwrightArgs}}",
      "env": {
        "API_HOST": "{{apiHost}}"
      }
    }
  }
}

Variables are defined in universal-ai-config.config.ts:

export default defineConfig({
  variables: {
    playwrightArgs: ["-y", "@playwright/mcp@latest"],
    apiHost: "api.example.com",
  },
});

For environment-specific values, use universal-ai-config.overrides.ts (gitignored):

export default defineConfig({
  variables: {
    playwrightArgs: [
      "-y",
      "@playwright/mcp@latest",
      "--headless",
      "--executable-path",
      "/usr/bin/chromium",
    ],
  },
});

Important: {{varName}} is for uac config variables. Use ${ENV_VAR} for runtime environment variable references that should be passed through to the generated output unchanged.

Per-Target Overrides

Any server field can have per-target values using the same override syntax as hooks:

{
  "mcpServers": {
    "my-server": {
      "command": {
        "default": "npx",
        "cursor": "node"
      },
      "args": {
        "default": ["-y", "@my/server"],
        "cursor": ["./mcp-server.js"]
      },
      "type": {
        "claude": "stdio",
        "copilot": "stdio"
      }
    }
  }
}

If command (and url) resolve to undefined for a target, the entire server is dropped for that target. This lets you define target-exclusive servers.

Copilot Inputs

For Copilot, you can include an inputs array for interactive secret prompts:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "${input:github-token}"
      }
    }
  },
  "inputs": [
    {
      "type": "promptString",
      "id": "github-token",
      "description": "GitHub Personal Access Token",
      "password": true
    }
  ]
}

The inputs array is only included in Copilot output — Claude and Cursor ignore it.

Output Paths

Generated MCP files are placed at root-relative paths (not inside the target's output directory):

Note: Cursor omits the type field from generated output (it infers transport from the presence of command vs url).

Example

{
  "mcpServers": {
    "github": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "${GITHUB_TOKEN}"
      }
    },
    "postgres": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres", "{{dbConnectionString}}"]
    }
  }
}