Best MCP servers for coding agents (by job)

"Best MCP servers" lists fail when they rank products instead of jobs. Coding agents need a small stack of specialized tools. Overlap wastes context; gaps force the model to invent.

This is a job map, not a leaderboard. Install what matches the work. Soft wording on third-party servers: verify install one-liners and auth on each vendor's docs the day you wire them.

Affiliation: BotKelp publishes this page and sells the verified scaffolding slot below. We do not claim the other slots.

Job map

1. Verified scaffolding — BotKelp

Use when the agent must compose a Next.js App Router greenfield tree from known IDs instead of inventing Tailwind, auth, and database wiring.

• Remote: https://www.botkelp.com/mcp
• Install example: claude mcp add --transport http botkelp https://www.botkelp.com/mcp
• Catalog: https://www.botkelp.com/catalog.json
• Always include nextjs-base; unknown IDs rejected; max components typically 12
• Payload: kind=data, instructions=false; require INTEGRITY.json verdict pass
• Knowledge tools search the catalog, not the open web
• Connection free; account tools use MCP OAuth 2.1 or Authorization: Bearer bk_live_… on the connection, never a tool argument
• Or pay per scaffold without an account via x402 (USDC on Base, 0.5 USDC; confirm /pay.json says payable before sending)

BotKelp should not be your docs browser or your GitHub client.

2. Live library docs — Context7

Use when the agent needs up-to-date library documentation instead of training memory.

• Typical tools: resolve-library-id, query-docs
• Remote often: https://mcp.context7.com/mcp
• Example: claude mcp add --transport http context7 https://mcp.context7.com/mcp
• Common shape: free tier plus API key for higher limits (often via CONTEXT7_API_KEY header) — confirm on Context7's site

Complementary to BotKelp, not a substitute either way. See /blog/botkelp-vs-context7.

3. Repo and PR operations — GitHub MCP

Use when the agent must open PRs, read checks, or triage issues against a real repository. Prefer the official or widely maintained GitHub MCP for your client; auth is usually a token with least privilege. We are not publishing a one-liner here as absolute — check current GitHub and client docs.

4. Browser and UI checks — Playwright-class MCP

Use when a passing build is not enough and you need navigation, screenshots, or basic UI assertions. Several Playwright-oriented MCP servers exist; pick one your client supports and keep credentials out of prompts.

5. Web research — Exa / Firecrawl-class MCP

Use for open-web discovery and extraction when catalog knowledge is the wrong corpus. These tools find pages; they do not stamp a coherent starter. See /blog/why-not-bootstrap-from-web-search.

Minimal stacks that work

Scaffold and ship Next features

1. BotKelp (compose) 2. Context7 (docs while coding) 3. GitHub MCP (PR)

Research-heavy spike

1. Research MCP (Exa/Firecrawl-class) 2. Context7 (library truth) 3. BotKelp only when you commit to a starter

UI bugbash

1. Playwright-class MCP 2. GitHub MCP 3. Docs MCP as needed

Anti-patterns

• One do-everything MCP and hope
• Using BotKelp knowledge tools as a general search engine
• Using web crawl output as package.json
• Stuffing API keys into chat instead of tool or header config
• Skipping integrity and build verification after generate

BotKelp quick start

claude mcp add --transport http botkelp https://www.botkelp.com/mcp

Human index: https://www.botkelp.com/for-agents · tools: https://www.botkelp.com/tools.json · security: https://www.botkelp.com/security.json