MCP

The Model Context Protocol (MCP) is an open standard for connecting AI agents to external tools and data sources. In 4RGE, MCP is one of the tool categories the tool broker exposes to agents, so MCP tools are used within the same controlled boundaries as everything else.

MCP lets agents query live systems — issue trackers, observability, databases, cloud APIs — instead of guessing from stale context in the repo.

How MCP fits in

1. Server provides tools

   An MCP server exposes tools (and sometimes read-only resources) with typed inputs and outputs.

2. 4RGE registers the server

   You install a plugin or add a custom server in Tools & MCP.

3. Broker surfaces tools to the agent

   MCP tools appear alongside repository, terminal, browser, and workspace tools.

4. Agent calls with approval

   The agent invokes MCP tools when relevant; sensitive calls pause for your approval.

• MCP servers provide tools (and sometimes read-only resources) that an agent can call.
• These tools are surfaced to agents through the tool broker alongside repository, terminal, browser, and workspace tools.
• Many ready-made MCP integrations are available through the plugin marketplace.

Transport types

4RGE supports the common MCP transport modes:

HTTP server example

{
  "my-service": {
    "type": "http",
    "url": "https://mcp.example.com/mcp"
  }
}

Hosted integrations (Linear, Datadog, Clerk, and many marketplace plugins) typically use HTTP with OAuth — 4RGE prompts you to authenticate when you connect the server.

stdio server example

{
  "my-local-server": {
    "type": "stdio",
    "command": "uvx",
    "args": ["some-mcp-package", "--transport", "stdio"]
  }
}

For plugin-bundled MCP, paths can use ${PLUGIN_ROOT} in command, args, cwd, and string env values — 4RGE Desktop expands the variable when registering plugin MCP.

Plugin MCP vs. custom servers

Plugin manifests use a flat map of server id → definition in mcp.json:

{
  "linear": {
    "type": "http",
    "url": "https://mcp.linear.app/mcp"
  }
}

Do not wrap the entire file in a nested { "mcpServers": { ... } } object unless you unwrap it to the flat shape above.

What agents can do with MCP

Examples from common marketplace integrations:

Prompt explicitly when you want MCP used: "File a Linear issue" or "Check Datadog logs for checkout 5xx in the last hour."

Using MCP tools safely

• MCP tool calls are mediated by the tool broker and subject to approval like other tools.
• Scope credentials to least privilege — prefer read-only connections for inspection tasks.
• Disconnect servers you no longer use from Tools & MCP.
• On cloud agents, ensure egress policy allows the MCP host.

Plugin registry and distribution

Marketplace plugins are indexed from the monorepo registry and served to clients from registry.4rged.ai in production (http://localhost:3005 locally). When you install a plugin from the in-app Marketplace, 4RGE downloads the versioned bundle, validates .4rge-plugin/plugin.json, and expands ${PLUGIN_ROOT} in stdio MCP definitions before registering servers with the broker. Custom servers you add under Tools & MCP skip the registry but use the same HTTP or stdio transport shapes documented above.

Connecting an MCP server

1. Install or configure

   Install a plugin from the marketplace, or add a custom server in Tools & MCP.

2. Authenticate

   Complete OAuth or enter API credentials as prompted. Credentials stay out of chat prompts.

3. Verify tools

   Confirm the server shows connected and tools are listed in Tools & MCP.

4. Prompt the agent

   Reference the integration in your request so the agent selects the right MCP tools.