Skip to content

AGENTS.md

Latest commit

 

History

History
105 lines (79 loc) · 3.91 KB

AGENTS.md

File metadata and controls

105 lines (79 loc) · 3.91 KB
Raw

AI Agent Instructions for Mapbox MCP Documentation Server

Note: If you're using Claude Code specifically, see CLAUDE.md instead. This file is for general AI coding assistants (Cursor, Continue, Aider, Copilot, etc.).

What This Project Does

This is an MCP (Model Context Protocol) server that gives AI assistants access to Mapbox documentation and reference materials. No Mapbox access token is required for most tools — it is a documentation-only server.

Tech Stack

  • Runtime: Node.js 22+
  • Language: TypeScript (strict mode)
  • Build: tshy (dual ESM/CJS output)
  • Testing: Vitest
  • Package Manager: npm

Project Structure

src/
├── index.ts                    # MCP server entry point
├── config/toolConfig.ts        # CLI argument parser
├── constants/                  # Static data (API endpoint definitions)
├── resources/                  # MCP resource implementations
│   ├── BaseResource.ts         # Abstract base class
│   ├── resourceRegistry.ts     # Resource registration
│   ├── utils/docParser.ts      # Shared doc parsing utilities
│   └── */                      # Individual resource implementations
├── tools/                      # MCP tool implementations
│   ├── BaseTool.ts             # Abstract base class
│   ├── toolRegistry.ts         # Tool registration
│   └── */                      # Individual tool implementations
└── utils/
    ├── httpPipeline.ts         # HTTP pipeline (User-Agent, caching, retry)
    ├── types.ts                # Shared types
    └── versionUtils.ts

test/                           # Mirrors src/ structure

Critical Patterns

Tool Architecture

  • All tools extend BaseTool<InputSchema, OutputSchema> (not MapboxApiBasedTool — that is in a different repo)
  • execute(input) receives only the validated input — no accessToken or context parameters
  • If a tool needs a Mapbox token, it must be a field in the input schema
  • Register tools in src/tools/toolRegistry.ts
export class MyTool extends BaseTool<typeof MyInputSchema> {
  readonly name = 'my_tool';

  constructor({ httpRequest }: { httpRequest: HttpRequest }) {
    super({ inputSchema: MyInputSchema });
    this.httpRequest = httpRequest;
  }

  protected async execute(
    input: z.infer<typeof MyInputSchema>
  ): Promise<CallToolResult> { ... }
}

HTTP Requests

  • Never patch global.fetch — use the injected httpRequest function
  • Pass httpRequest via constructor so tools can be tested without network access
  • The shared httpRequest from src/utils/httpPipeline.ts adds User-Agent, 1-hour caching, and retry

Testing

  • Use Vitest exclusively
  • Mock all HTTP calls — no real network requests in tests
  • Place tests in test/ mirroring the src/ structure

Essential Commands

npm install              # Install dependencies
npm test                 # Run tests
npm run build            # Compile TypeScript
npm run lint             # ESLint
npm run format           # Prettier check
npm run inspect:build    # Test with MCP Inspector

Common Pitfalls

  1. Don't use MapboxApiBasedTool — it doesn't exist here, use BaseTool
  2. Don't patch global.fetch — use the httpRequest pipeline
  3. Don't make real network calls in tests — mock httpRequest with a stub
  4. Don't read the access token from an env variable — accept it as an input field if needed
  5. Don't commit without updating CHANGELOG.md — add an entry under Unreleased

Pull Requests

  • Update CHANGELOG.md under the Unreleased section
  • All tests must pass: npm test
  • Lint and format must pass: npm run lint && npm run format

Further Reading