Claude Code (HTTP)

Recommended. Claude Code has first-class support for MCP over HTTP. No shim, no extra process.

Setup

After the resource boots, the FiveM console prints a copy-paste-ready JSON block. Paste it under mcpServers in one of these files:

• ~/.claude.json — user-global (applies to every project)
• <project>/.mcp.json — project-scoped

{
  "mcpServers": {
    "agent_api": {
      "type": "http",
      "url": "http://127.0.0.1:30120/agent_api/mcp",
      "headers": {
        "x-agent-token": "PASTE_YOUR_TOKEN_HERE"
      }
    }
  }
}

Restart any open Claude Code session. The MCP server appears in /mcp with 36+ tools and one prompt.

Verify

claude
> /mcp

You should see:

agent_api · 36 tools, 1 prompt

If you see fewer than 36 tools, check that the plugins succeeded:

> list_plugins

Using the scaffold prompt

> /agent_api:scaffold-fivem-resource

This injects the grill workflow into the conversation. Claude will refuse to write any files until you've answered every question.

You can also trigger the same workflow by intent — just say "create a new fivem resource" and Claude will auto-call the scaffold_fivem_resource_workflow tool because its description matches the intent.

Notes

• Token rotation: delete dist/.agent_token, restart agent_api, paste the new token into your config.
• The HTTP server binds to 127.0.0.1 only. To reach agent_api from a different machine, either change the bind (advanced) or SSH-tunnel 30120 to your dev machine.
• Body limit is 5 MB. Tool calls returning more than that are extremely rare.