The best MCP servers for Claude Code (and the config that actually works)

AI Prime Tech — Wed, 12 Aug 2026 23:11:58 +0000

Model Context Protocol turns Claude Code from a code generator into something that can read your repo,
query a database, drive a browser and file a PR. The catch is that most "MCP setup" posts stop at
filesystem and never mention the traps. Here's a config I actually run, plus the gotchas.

A `.mcp.json` that earns its place

Drop this in your repo root; Claude Code auto-detects it and asks you to trust it once.

{
  "mcpServers": {
    "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "${workspaceFolder}"] },
    "github":     { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_TOKEN": "..." } },
    "git":        { "command": "uvx", "args": ["mcp-server-git", "--repository", "${workspaceFolder}"] },
    "fetch":      { "command": "uvx", "args": ["mcp-server-fetch"] },
    "playwright": { "command": "npx", "args": ["-y", "@playwright/mcp@latest"] }
  }
}

(Full pack with a postgres server and per-tool notes:
https://github.com/iskandaryv/anthropic-gateway-examples/tree/main/mcp)

The gotchas that cost real time

The trust prompt is a feature. Claude Code refuses to load project MCP servers until you approve them — per project. That's the security boundary, not a bug.
Give postgres a read-only role. An MCP server runs whatever SQL the model emits. A read-only role makes a hallucinated DROP physically impossible.
Scope the GitHub token. repo + read:org is enough. Don't hand it an all-scopes classic PAT.
MCP output is context. A 200-row query result or a full page fetch gets re-sent on the next turn. Big tool outputs inflate token usage fast — the thing that makes agentic sessions expensive.

It works the same through a gateway

MCP wiring is independent of where model calls go. Whether Claude Code talks to the official API or an
Anthropic-compatible gateway, the servers
are configured identically — the gateway only changes ANTHROPIC_BASE_URL, not .mcp.json.

Takeaways

Start with filesystem + git + github; add browser/db only when a task needs them.
Treat every server as an untrusted executor — least-privilege tokens, read-only DB roles.
Watch token usage: tool results are the hidden cost of an agentic loop.

Disclosure: I run ClaudeAPIKey.dev, an Anthropic-compatible API
gateway — mentioned above only because MCP config is unaffected by it. "Claude" is a trademark of
Anthropic; we are not affiliated with Anthropic.

Cutting Claude Code token costs: where the tokens actually go

AI Prime Tech — Wed, 12 Aug 2026 23:10:54 +0000

If you've run Claude Code for a full working day, you've seen the bill. Agentic loops are
token-hungry in a way normal chat never is — every file read, every tool result, every diff lands in
context and gets re-sent on the next turn. Here's where the tokens go and how to stop the leak.

Output tokens are the expensive ones

On every frontier model, output tokens cost 4–5× input. A verbose model that "thinks out loud" in the
final answer burns money faster than one that's concise. Cap max_tokens deliberately, and prefer a
model that answers tightly for routine work.

Quick way to see the shape of a task's cost before you run it — this reads a live model catalog and
prints per-model cost for the token counts you expect:

python3 cost_estimate.py --in 20000 --out 2000
# MODEL                CTX          COST (USD)
# Claude Haiku 4.5     200K            $0.0120
# Claude Sonnet 4.6    200K/1M         $0.0270
# Claude Opus 4.8      200K            $0.0900

(The script is in an open repo: https://github.com/iskandaryv/anthropic-gateway-examples)

The four levers that actually move the bill

Route by task. Classification, extraction and routing don't need Opus — Haiku does them at a fraction of the price. Reserve the flagship for reasoning-heavy work. Full comparison: all Claude models compared.
Prompt caching. Repeated system prompts and stable prefixes are re-read from cache at ~10% of input price on supported endpoints. For agentic tools this is the single biggest lever.
Trim context each turn. Don't resend the whole transcript; keep the last few turns plus a running summary. Unbounded context growth is what turns a $2 session into a $20 one.
Cap the loop. Set a hard step limit on autonomous runs so a stuck agent can't spin forever.

The gotcha almost everyone hits

If ANTHROPIC_API_KEY is set in your shell, Claude Code bills that key at API rates and silently
ignores your Pro/Max subscription. Unset it when you mean to use the subscription. (More setup
traps in the Claude Code API costs guide.)

Takeaways

Watch output tokens, not prompt length — that's where the money is.
Route cheap-first, escalate only when quality demands it.
Turn on prompt caching for anything with a stable prefix.
Estimate before you run; measure tokens in/out after.

Disclosure: I run ClaudeAPIKey.dev, an Anthropic-compatible API
gateway, so read the pricing notes as coming from an interested party. The token-accounting advice
above applies identically on the official API. "Claude" is a trademark of Anthropic; we are not
affiliated with Anthropic.

DEV Community: AI Prime Tech

The best MCP servers for Claude Code (and the config that actually works)

A .mcp.json that earns its place

The gotchas that cost real time

It works the same through a gateway

Takeaways

Cutting Claude Code token costs: where the tokens actually go

Output tokens are the expensive ones

The four levers that actually move the bill

The gotcha almost everyone hits

Takeaways

A `.mcp.json` that earns its place