Best MCP Servers for Roo Code (2026)

The best MCP servers for Roo Code are the few that earn their tool budget: an official GitHub server, a docs/context server that cuts hallucinations, and one data server for whatever you actually query. Everything else is a liability, because tool selection degrades past roughly 40 active tools and the average server ships around 12. So the real question isn't "which servers exist" — it's which three or four are worth the slots.

Roo Code has one advantage here that Cursor doesn't: you can flip a server's disabled flag and Roo strips its tools out of the system prompt entirely, cutting token usage. That makes an over-loaded config cheaper to recover from — but it doesn't excuse installing servers you won't use. This is the opinionated shortlist, with Roo's config path, honest trade-offs, and what to skip.

Where Roo Code reads MCP config (do this first)

Roo Code reads MCP servers from two places: a global mcp_settings.json (open it via the MCP pane's settings icon → Edit Global MCP) and a per-project .roo/mcp.json in your repo root. Project config wins when a name collides, so commit .roo/mcp.json to share a team setup and keep secrets in the global file.

Both files use the same mcpServers object, and Roo supports STDIO for local servers plus Streamable HTTP and SSE for remote ones. A minimal entry:

{
  "mcpServers": {
    "github": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "ghcr.io/github/github-mcp-server"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." },
      "disabled": false
    }
  }
}

Save and Roo reloads automatically — check the MCP pane for a green status dot. Use alwaysAllow: ["tool_name"] to skip approval prompts for tools you trust, and disabled: true to park a server without deleting it. Full walkthrough in how to add an MCP server.

The tool budget: why three servers, not thirty

Assistants like Roo Code work best under about 40 active MCP tools, and the average server exposes roughly 12. That means you hit the wall around four servers, not fifty. Past the budget, the model faces too many near-identical tools and starts calling the wrong one — which reads like the model got dumber but is really a config problem.

Treat slots as scarce. Every server below does one job the others don't, and most run locally over STDIO (as ≈90% of MCP servers do), which keeps latency low and your code off someone else's logs. Roo's disabled toggle helps you stay under the ceiling: keep a rarely-used server installed but off, and enable it only for the session that needs it. The arithmetic is the same one I ran in the tool-limit math post.

The shortlist: MCP servers worth adding to Roo Code

Start with a code server, a context server, and one data server. Add a fourth only if you'll use it weekly.

Server	Best for	Transport	Add it if
GitHub MCP Server	Repos, issues, PRs, Actions	STDIO (Docker) or remote	You do PR/issue work from Roo
GitMCP	Live docs for any repo	Remote, free	You want fewer hallucinated APIs
Octocode	Deep code research (FS + LSP)	STDIO	You investigate large codebases
Package Registry MCP	NPM/PyPI/crates versions + advisories	STDIO	You get burned by stale package data
Anyquery	SQL over files, DBs, 40+ apps	STDIO	You query structured data often
Graphlit MCP Server	RAG over Slack/Drive/Jira/Linear	Remote	Your context lives across SaaS tools

The one everyone should have: GitHub

The GitHub MCP Server is GitHub's official server, and "official" earns its place — it tracks the API and won't quietly rot the way abandoned community forks do. It covers repos, issues, PRs, and Actions, running locally in Docker or via GitHub's hosted remote. Trade-off: it's tool-heavy, so it eats a real chunk of your budget. If you only read issues, scope the token tightly and lean on Roo's alwaysAllow for the two or three tools you use, not all of them.

Cut hallucinations: GitMCP

GitMCP turns any GitHub repo into a live documentation source — the cheapest fix for the model inventing method signatures. It's a free remote server, so there's no install; add the Streamable HTTP URL and go. I reach for this over pasting a whole SDK into context: it grounds Roo on the actual repo docs instead of the model's stale memory. Skip it only if you already run a broader RAG server.

Go deep on code: Octocode

Octocode is the pick when "search" isn't enough and you need to trace how code actually works. It wires GitHub, the local filesystem, and LSP intelligence into Roo, so the assistant follows definitions and references like a staff engineer reading unfamiliar code. It overlaps with the GitHub server, so run both only if you do heavy investigation — otherwise this replaces it.

Stop trusting stale package data: Package Registry

The Package Registry MCP Server searches NPM, PyPI, crates.io, NuGet, and Go registries — plus GitHub Security Advisories — for current versions. The model's training data is months old; this makes "what's the latest, and is it vulnerable" a live lookup instead of a guess. It's small and cheap on tools, so it's an easy fourth slot.

One data server, chosen by your workflow

Pick exactly one. Anyquery runs SQL over files, databases, and 40+ apps (GitHub, Notion, Chrome, Todoist) and exposes them to Roo over MCP — reach for it if you think in queries. Graphlit MCP Server is the RAG option: ingest and search across Slack, Discord, Google Drive, Jira, and Linear when your real context is scattered across SaaS tools. Running both is how you blow the budget.

What to skip

Skip the maximalist bundles. Kitchen-sink servers that expose 30+ tools each will consume your entire budget in two installs and tank tool selection — the exact failure the tool-limit math predicts.

Skip duplicate coverage. Running GitHub, Octocode, and a third code server means three overlapping tool sets competing for the same call. Pick the one that matches how you work, and disabled: true the rest.

Skip community forks of things that ship an official server. When an official option exists, the maintenance and security story is simply better — check the official-vs-community badge before you add anything. For deeper comparisons, the best MCP servers list and best servers for coding agents rank options by category.

My default Roo Code stack

For most engineers: GitHub (official) + GitMCP (docs) + Package Registry (versions) + one data server. That's four servers, comfortably under the ~40-tool ceiling, each doing a distinct job. Swap Octocode in for the GitHub server if your day is code archaeology rather than PR management. Add nothing else until you feel a real gap — and because Roo lets you disabled-toggle servers per project, keep the extras installed but off rather than piling them all into one bloated global config.