What MCP actually is.
The Model Context Protocol is a small open spec that lets AI tools call external services in a uniform way. A server exposes tools (callable functions) and resources (readable URLs). Your agent's host (Cursor, Claude Code, Codex CLI, Aider, OpenCode) connects, sees the tool list, and decides when to call which.
The win is composability. Install the Playwright MCP and every agent on your machine gains the ability to drive a browser. The cost is honest: every installed server adds its tool schema to every turn. That schema is real tokens — typically 400 to 1200 per server, sometimes more.
- Tools = verbs.
browser_click,db_query,issue_create. - Resources = nouns. A Postgres MCP exposes table schemas as resources the agent can read on demand.
- STDIO transport for local servers (filesystem, local DB, Playwright). HTTP/SSE transport for SaaS-hosted ones (Linear, Sentry).
The seven worth installing.
Ranked by frequency of "I'm glad I have this" moments per week for a working web developer. Token costs are approximate, measured against a Claude or GPT-5 schema dump:
| Server | Use it for | Cost / turn | Verdict |
|---|---|---|---|
| Playwright MCP | Driving the browser to verify a UI change. Capturing screenshots for review. Reading the console for client-side errors. | ~900 tok | Install |
| Filesystem MCP | Letting an agent read files outside the current workspace (design assets, shared notes). Most editors handle in-workspace files natively. | ~500 tok | Maybe |
| Postgres / Drizzle MCP | Inspecting the dev DB schema, running read-only queries, understanding what the data actually looks like before writing a query. | ~800 tok | Install |
| GitHub MCP | Reading issues, opening PRs, leaving review comments, fetching the failing CI log for the run that just blew up. | ~1100 tok | Install |
| Linear / Jira MCP | Pulling the ticket the work is for, updating status when a PR opens, linking commits back. Especially useful with background agents. | ~700 tok | Install |
| Sentry / AppSignal MCP | "Why did production 500 yesterday?" The agent fetches the exception, walks the backtrace, proposes a fix grounded in real evidence. | ~600 tok | Install |
| Figma MCP | Translating a design frame into accurate spacing, colors, and component structure. Lifesaver if you build from Figma; dead weight if you don't. | ~850 tok | Designer-dependent |
| Slack MCP | Looks useful, almost never is. Agents posting to Slack is more chaos than value 95% of the time. | ~700 tok | Skip |
| Brave / Web Search MCP | Tempting, but most modern agents have built-in web search. Don't double-pay. | ~400 tok | Skip |
A reasonable default for a working web developer: Playwright + Postgres + GitHub + Linear + Sentry. That's about 4,100 tokens of overhead per turn. At Sonnet input rates, roughly $0.012 per 1,000 turns. Worth it.
Install snippets.
Three syntaxes, same idea. Pick your host:
{
"mcpServers": {
"playwright": {
"command": "npx",
"args": ["-y", "@playwright/mcp@latest"]
},
"postgres": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-postgres", "postgres://localhost/dev"]
},
"github": {
"url": "https://api.githubcopilot.com/mcp/",
"headers": { "Authorization": "Bearer ${GITHUB_TOKEN}" }
}
}
}claude mcp add playwright npx -y @playwright/mcp@latest claude mcp add postgres npx -y @modelcontextprotocol/server-postgres postgres://localhost/dev claude mcp add --transport http github https://api.githubcopilot.com/mcp/ claude mcp list
[mcp_servers.playwright] command = "npx" args = ["-y", "@playwright/mcp@latest"] [mcp_servers.postgres] command = "npx" args = ["-y", "@modelcontextprotocol/server-postgres", "postgres://localhost/dev"]
Your agent will, at some point, decide a DELETE is the cleanest way to test a hypothesis. Connect MCP to a read-only role — Postgres makes this one CREATE ROLE … LOGIN away. Migrations and writes go through your normal app code, not the agent's loop.
Live: build your MCP stack.
Toggle the servers you want. The widget estimates context tax, monthly cost at typical usage, and warns when you're past the "agent picks wrong tools" cliff. All math runs in your browser.
Aim for the green zone — under about 4,500 tokens of overhead. Past that you're paying for tools the agent won't pick anyway, because the choice space is too wide.
Common pitfalls.
The MCP server directory is full of impressive demos. Most of them solve someone else's workflow. Install on a "earned it this week" basis: if you didn't use a server's tools in the last seven days, uninstall it.
Your Postgres MCP pointing at a prod read replica is one prompt-injection away from leaking the user table. Local dev DB only. Production data goes through your normal observability tools, not MCP.
STDIO servers fork on every editor reload. Twelve installed servers means twelve spawned Node processes lurking in the background, each holding a megabyte or two of memory. ps aux | grep mcp occasionally to keep yourself honest.