Best MCP Servers for VS Code (2026)

The best MCP servers for VS Code in 2026 are GitHub MCP Server, GitMCP, Package Registry MCP, Octocode, Anyquery, and Graphlit. VS Code's Copilot agent mode speaks MCP natively, so the only real question is which servers pull their weight against a limited tool budget — and which ones quietly wreck it. This guide is the short list, the config, and the honest cuts.

Most of these run locally over stdio, which matches the ecosystem: roughly 90% of MCP servers run locally over stdio rather than as remote HTTP services. That has real consequences for VS Code — local servers need Docker or a runtime on your machine, remote ones need OAuth and nothing else.

How VS Code loads MCP servers (and why the budget matters)

VS Code loads MCP servers from a .vscode/mcp.json file in your workspace (or your user settings), and every enabled server dumps its tools into one shared list the model sees. That shared list is the catch. Language models degrade as the tool count climbs, and most clients ship a soft ceiling around a 40-tool budget before selection accuracy falls off a cliff.

So adding servers is not free. A single server can register 20+ tools; two greedy ones and you have blown past the budget before writing a line of code. The tool-limit math is the same in VS Code as it is in Cursor: fewer, sharper servers beat a pile of overlapping ones.

The practical rule: install servers per-workspace, not globally, and disable the ones you are not using this week. VS Code lets you toggle individual servers and even individual tools — use it.

Here is the minimal shape of that file:

{
  "servers": {
    "github": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "GITHUB_PERSONAL_ACCESS_TOKEN",
        "ghcr.io/github/github-mcp-server"
      ],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "${input:github_pat}" }
    }
  }
}

If you would rather not hand-write JSON, each server's page in the best MCP servers catalog carries its ready-to-paste config block.

GitHub MCP Server — the one non-negotiable

GitHub MCP Server is the first server most VS Code users should add, because agent mode without repo, issue, and PR access is half a tool. It is GitHub's official server (not a community port), and it replaces the deprecated @modelcontextprotocol/server-github reference package.

You get two paths. Run it locally with the ghcr.io/github/github-mcp-server Docker image and a personal access token — that is the config above. Or connect to GitHub's hosted remote server at https://api.githubcopilot.com/mcp/ over OAuth, which skips Docker entirely and is the lower-friction choice on a fresh machine.

The trade-off is tool count: it exposes a lot of them, spanning repos, issues, PRs, and Actions. Turn off the toolsets you do not need (Actions, for instance) so it does not eat your whole budget alone.

GitMCP and Package Registry — kill hallucinated APIs

If your agent keeps inventing function signatures and outdated package versions, these two fix it directly. GitMCP is a free remote server that turns any GitHub repo into a live documentation source, so the model reads the actual current docs instead of guessing from training data. It is remote and hosted — no install, just a URL — which makes it the cheapest possible add to any workspace.

Package Registry MCP covers the other half: it searches NPM, crates.io, NuGet, PyPI, and Go registries for current versions, plus GitHub Security Advisories. When the agent proposes a dependency, it can check that the version exists and is not flagged, instead of pinning something from 2023.

Both are small, focused servers — exactly the profile that survives the tool budget. Add them early.

Octocode and Anyquery — code research and local data

For deeper code work, Octocode researches code the way a staff engineer does: it wires GitHub, your local filesystem, and LSP intelligence into the assistant, so it can trace how a symbol is actually used across a codebase rather than pattern-matching on names. It overlaps with GitHub MCP Server on the GitHub side, so pick one as your primary and disable the other's redundant tools.

Anyquery is the wildcard: it runs SQL over files, databases, and 40+ apps — GitHub, Notion, Chrome, Todoist — and exposes all of it to the model over MCP. In VS Code this is how you let the agent query a local SQLite file or a CSV without you writing a script. It is broad, so scope it to what you need per workspace.

Graphlit — when you need RAG over your team's data

Graphlit is the pick when the context you need lives outside the repo — in Slack, Discord, Google Drive, Jira, or Linear. It ingests those sources, indexes them, and serves search and RAG back to the agent, so "why did we decide X" can be answered from the actual thread instead of your memory.

This is not an everyday server. It is worth it on a team codebase where decisions and tickets carry as much weight as the code. Skip it for solo or greenfield work — the ingestion overhead is not justified.

What to install, and what to skip

Here is the short version, ranked by how often it earns its slot:

Server	Transport	Add it when	Skip it when
GitHub MCP Server	stdio (Docker) or remote	Always — you work with GitHub	Never; disable unused toolsets instead
GitMCP	remote (hosted)	Agent hallucinates APIs	You work fully offline
Package Registry	stdio	You add dependencies often	Config-frozen projects
Octocode	stdio	Deep cross-repo code research	GitHub MCP already covers you
Anyquery	stdio	You query local files/DBs	Pure code editing
Graphlit	stdio	Team context lives in Slack/Jira	Solo or greenfield

What to skip in general: any server whose job overlaps one you already run, and anything with a huge undifferentiated tool count you cannot narrow. Two focused servers beat five broad ones every time in VS Code. For a wider catalog, browse the best MCP servers list or the picks tuned for coding agents.