Best MCP Servers for Zed (2026)

The best MCP servers for Zed in 2026 are a small set: GitHub MCP Server for repo work, GitMCP to kill doc hallucinations, Package Registry MCP for version-accurate deps, and Octocode for cross-repo code research. Zed wires MCP servers in as "context servers," and the agent has a finite tool budget, so adding everything is the wrong move. Below is what to add, the config, and what to skip.

This guide assumes Zed's Agent Panel. If you're new to the protocol, browse what MCP servers can do by capability; if you want the cross-editor picture, see the best MCP servers list.

Where Zed keeps MCP config

Zed defines MCP servers under a context_servers block in settings.json (open it with cmd-,). Roughly 90% of servers run locally over stdio, meaning Zed spawns a process and talks to it over standard input/output — so most entries are just a command plus args and env.

A minimal stdio entry looks like this:

{
  "context_servers": {
    "github": {
      "source": "custom",
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "GITHUB_PERSONAL_ACCESS_TOKEN", "ghcr.io/github/github-mcp-server"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    }
  }
}

Remote servers use a url instead of a command. After saving, reopen the Agent Panel and confirm the server shows its tools. If it hangs, that's almost always a bad command path or a missing env var, so double-check the exact command and every entry in env before looking anywhere else.

Watch the tool budget before you add anything

Each server exposes tools, and every tool description is injected into the model's context on every request. Practical clients degrade past roughly 40 tools — the model starts mis-picking or ignoring tools once the list gets long. This is the same math that bites Cursor users; the tool-limit math applies to Zed too.

So treat the budget as real. Three or four focused servers beat eight kitchen-sink ones. If a server dumps 30 tools and you use two of them, it's costing you accuracy on everything else.

GitHub work: the official server

For anything touching repos, issues, PRs, or Actions, add the GitHub MCP Server — GitHub's official one. It runs as a local Docker container over stdio, or you can point Zed at GitHub's hosted remote endpoint if you'd rather not run Docker.

The catch: it's tool-heavy. Use its toolset flags (or the remote's scoped configuration) to enable only the groups you need — repos, issues, pull_requests — instead of the full surface. That keeps you inside the budget above. Prefer official over community forks here; auth and rate-limit handling are the parts you don't want a hobby fork guessing at.

Cut hallucinations: docs and package facts

Two servers earn their slot by feeding the model current facts instead of its training-cutoff memory.

GitMCP is a free remote server that turns any GitHub repo into a live documentation source. Point it at a library's repo and the agent reads that project's actual docs and code, which is the cheapest fix for "the model confidently used an API that doesn't exist." Being remote, it needs no local process — a rare case where remote is the right default.

Package Registry MCP answers "what's the current version and does it have a known advisory" across NPM, crates.io, NuGet, PyPI, and Go, plus GitHub Security Advisories. Pair it with GitMCP and the agent stops inventing version numbers.

Code research and data access

For reasoning across a codebase, Octocode wires GitHub, the local filesystem, and LSP intelligence into the agent so it can research code the way a staff engineer would — trace a symbol, read the definition, check how a repo uses it. It overlaps with Zed's own project awareness, so add it when you work across many repos, not for a single small project.

If your problem is data rather than code, two options with different shapes:

• Anyquery lets the agent run SQL over files, databases, and 40+ apps (GitHub, Notion, Chrome, Todoist). Add it when "query my stuff" is the job.
• Graphlit MCP Server ingests and does RAG over Slack, Google Drive, Jira, Linear, and more. It's heavier — a hosted platform, not a local script — so justify it before adding.

What to skip, and the picks table

Skip anything that duplicates Zed's built-ins: file read/write, terminal, and basic project search are native, so a generic "filesystem" server is wasted budget. Skip broad multi-tool aggregators unless you actually use the breadth — they're the fastest way to blow past 40 tools. And skip community rebuilds of official servers when the official one exists.

Server	Best for	Transport	Add it when
GitHub MCP Server	Repos, issues, PRs, Actions	Local Docker / remote	You do real GitHub work (scope the toolsets)
GitMCP	Live library docs	Remote	The model hallucinates APIs
Package Registry MCP	Version + advisory checks	Local stdio	You touch dependencies
Octocode	Cross-repo code research	Local stdio	You work across many repos
Anyquery	SQL over apps & files	Local stdio	Your task is querying data
Graphlit MCP Server	RAG over SaaS data	Hosted remote	You need org-wide knowledge search

Start with GitHub, GitMCP, and Package Registry. Add Octocode or a data server only when a real task demands it. For the client-agnostic version of these picks, see best MCP servers for coding agents.