The Best MCP Servers for VS Code in 2026

GitHub MCP server

This is GitHub's own MCP server, and the one I tell everyone to set up first. It exposes repositories, issues, pull requests, Actions runs, and code scanning alerts as tools Copilot can call directly. Instead of tabbing out to github.com to see why a workflow failed, you ask in the chat and it reads the run logs, then proposes a fix in the same thread. It shines at triage ("summarize the open issues labeled bug and group them"), PR review, and chasing down a red CI run.

Pricing: free. The remote endpoint lives at https://api.githubcopilot.com/mcp/ and uses one-click OAuth, so there is nothing to install and no personal access token on disk.

The standout is how little setup it needs. OAuth means the server only gets the scopes you approve, and you can append /readonly to the URL to strip out every write tool (per GitHub's remote-server docs). The --toolsets flag scopes it further so the agent is not drowning in 80 tools.

The catch: with every toolset on it is a lot of surface area, and a confused agent can open issues or push comments you did not want. Start read-only and turn on write toolsets deliberately.

Context7

Context7, built by Upstash, solves the single most maddening failure in AI coding: the model writing code against a library version that no longer exists. Training data goes stale, so Copilot will happily call a deprecated API or invent a function signature. Context7 fetches current, version-specific documentation for thousands of libraries at query time and injects it straight into the agent's context. If you live in Next.js, Tailwind, Prisma, or any SDK that ships breaking changes, this is the difference between code that runs and code that hallucinates an import.

Pricing: free. Run it with no API key on basic rate limits, or grab a free key from the Context7 dashboard for higher limits (their pricing confirms the Free plan is $0). Most solo developers never hit the free ceiling.

The standout is the workflow. Add "use context7" to a prompt, the server resolves the library, pulls the docs, and hands the agent accurate reference material. No copy-pasting docs into chat.

The catch: it only helps when the library is indexed, and coverage for very new or niche packages can lag. It also adds latency on the first call. Minor costs for what you get back.

Playwright MCP

Microsoft maintains the Playwright MCP server, which hands Copilot a real browser. The clever part is that it does not rely on screenshots. It exposes the page as a structured accessibility snapshot, a text representation where every interactive element gets a reference ID. The agent reads that, clicks the right element, fills forms, and navigates without burning tokens on vision models. Ask it to "open the staging login page, sign in with the test account, and tell me why the dashboard is empty" and it will drive the browser to find out.

Pricing: free and open source, shipped as both an npm package and a Docker image. Install it with npx @playwright/mcp and point your mcp.json at it.

The standout is reliability. Because it works off the accessibility tree instead of pixel-matching, it is far less brittle than screenshot-driven browser agents, and it still handles network mocking and screenshots when you need them.

The catch: it controls a live browser, so on a real site it can submit forms for real. Keep it pointed at local or staging, and be deliberate about production URLs. The first run also downloads browser binaries.

Supabase MCP

The official Supabase MCP server lets Copilot manage your backend: query the Postgres database, inspect schema, run migrations, and deploy edge functions. When the agent knows your actual table structure, the SQL it writes stops being guesswork and stops inventing column names. It is also handy for one-off data questions you would otherwise write a query for by hand.

Pricing: free to use; it runs against your existing project, so you pay only for the plan you already have. Setup runs through the hosted endpoint or npx supabase-mcp@latest.

The standout is the read-only safety valve. Append ?read_only=true and every mutating tool gets disabled, including migrations and edge-function deploys (Supabase documents this in their MCP guide). For anything touching production, that should be your default.

The catch: pointed at a writable production database, an agent that misreads your intent can run destructive SQL. Use read-only for prod and never hand it service-role access casually.

Sentry MCP

The Sentry MCP server connects Copilot to your error tracking. Instead of switching to the Sentry dashboard, copying a stack trace, and pasting it back, the agent pulls the issue details, searches events, and reads the exact file and line where things broke. The loop of "a user reported a bug, find it, fix it" collapses into a single chat session.

Pricing: free server that runs against your existing Sentry account, with a hosted remote endpoint at mcp.sentry.dev using OAuth 2.0, so there is no long-lived token on disk.

The standout is Seer integration. The server can invoke Sentry's AI agent for root-cause analysis, returning code-level explanations, file locations, and concrete fixes. I have watched it go from a vague exception to a one-line patch suggestion in a single prompt.

The catch: Seer's deeper analysis is a paid Sentry feature, so the most useful part depends on your plan tier. The basic issue-reading tools work on any account.

If you are wiring up a stack like this, a focused course beats poking at config files for an afternoon. Our Dupple X program walks through building real AI-assisted dev workflows end to end, including MCP setup. Start a yearly trial.

Figma Dev Mode MCP

Figma's official Dev Mode MCP server reads the live structure of your design files: the node tree, layout constraints, variables, and tokens, then exposes them to the agent. The result is Copilot generating component code that matches the spacing, colors, and tokens a designer set, instead of approximating from a screenshot. The closer your code needs to match the design, the more this earns its keep.

Pricing: free during the current beta, but it requires a Dev or Full seat on a paid Figma plan (Professional starts around $16 per editor per month billed annually). Starter accounts are capped at roughly 6 calls a month (Figma's MCP guide spells out the seat requirements).

The standout is fidelity. Because it reads the real file structure rather than guessing from pixels, the generated code respects the variables and tokens that make design systems hold together.

The catch: the seat requirement makes it the priciest entry here, and Figma has said it will eventually move to usage-based pricing. Treat the free beta as a window to decide if it is worth a seat.

Filesystem MCP

The official Filesystem MCP server from the Model Context Protocol team gives the agent controlled read and write access to files, scoped to directories you explicitly allow.

Best for anyone who wants the agent to work across files outside the current workspace, or who wants an auditable allowlist of what it can touch.

Pricing: free. Run it with npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/dir and pass the directories it is permitted to see. Directories are set by command-line argument or via Roots, so the server cannot reach anything outside what you granted.

The catch: VS Code's agent already has solid file editing built in, so for plain in-project work this is redundant. Reach for it when you need cross-directory access or a hard, declared boundary.

Brave Search MCP

The Brave Search MCP server gives the agent live web search through Brave's independent index, covering web, news, images, video, and local results. When Copilot needs current information, like a changelog or an error someone posted last week, this is how it finds it without you alt-tabbing to a browser.

Best for developers who want their agent to research in real time rather than rely on stale training data. Brave's index is its own, not a Google or Bing reseller, which is useful when debugging against the latest version of anything.

Pricing: this changed in early 2026. The old free 2,000-queries-a-month tier was retired; it is now pay-as-you-go at roughly $0.003 to $0.005 per query with a $5 prepaid credit, plus a Pro tier at $5 a month for 20,000 queries (Brave's API pricing has the details). Cheap, but no longer free.

The catch: it needs an API key and a card on file now, so it is the only entry here with a real per-use cost. For occasional lookups it is pennies; for an agent that searches constantly, watch the meter.