Awesome Free, Open-Source MCP Servers Worth Installing

Every week someone posts a "3,000 MCP servers!" directory and it is useless, because 90% of those entries are wrappers you will never install and can't verify. What you actually want is a short, honest shortlist: which free MCP servers are real, open-source, and worth wiring into your agent today.

This is that list. I build openseek, a TUI coding agent, so I connect and rip out MCP servers constantly. Below are the ones with public repos and permissive licenses that survive contact with real work — grouped by category, each with a one-line "add it when…" so you're not installing on vibes. If you're new to the concept, start with what is an MCP server and how to add an MCP server to Claude Code. For a broader ranked roundup, see the best MCP servers of 2026.

One note before the list: I'm deliberately not quoting star counts. They drift daily and half the numbers you see repeated online are stale. Every server here has a real, checkable repo — go look at the stars yourself.

The Official Reference Servers (start here)

The Model Context Protocol steering group maintains a small set of reference servers at modelcontextprotocol/servers, MIT-licensed. These are the canonical, genuinely free implementations — and the best code to read if you want to learn how a server is built.

• Filesystem — secure, scoped file operations with configurable access controls. Add it when you want the agent to read and edit files in a directory you explicitly allow-list, without handing it your whole disk.
• Git — read, search, and manipulate a local Git repo (diffs, logs, blame, commits). Add it when you want the agent reasoning over repository history, not just the working tree.
• Fetch — pulls a URL and converts it to clean Markdown for the model. Add it when you need lightweight "read this page" without a full browser.
• Memory — a knowledge-graph-based persistent memory store. Add it when you want the agent to remember entities and relationships across sessions.
• Sequential Thinking — a structured scratchpad for multi-step, revisable reasoning. Add it when a task needs the model to plan and backtrack out loud.
• Time — current time and timezone conversion. Add it when your agent keeps getting "what day is it" wrong (they do).
• Everything — a reference/test server that exercises every MCP feature (prompts, resources, tools). Add it when you're building your own client or server and need something to poke at.

Worth knowing: Anthropic archived its original batch of connector reference servers (Slack, Postgres, SQLite, Puppeteer, Brave Search, GitLab, and others) into servers-archived. Most of those capabilities are now maintained by the vendors themselves — the archive still works but isn't the place to look for active development.

Dev & Git

• GitHub's official server — github/github-mcp-server, rewritten in Go, MIT-licensed. Issues, PRs, code search, Actions runs, code-scanning alerts. Add it when you want the agent to triage issues or open PRs against a repo you have access to. This replaced the old archived reference server and is the one to use.
• GitMCP — idosal/git-mcp, a free, open-source remote server that points at any public GitHub project's docs and code. Add it when you want up-to-date grounding on a specific library without cloning it locally.
• Context7 — upstash/context7, run via npx -y @upstash/context7-mcp. Pulls version-specific library docs into the prompt. The MCP server is MIT and open source, though be honest with yourself: the backend crawling/parsing service behind it is proprietary. Add it when the model keeps hallucinating APIs from an old training snapshot.

Data & Databases

• DBHub — bytebase/dbhub, an open-source, zero-dependency server that speaks to PostgreSQL, MySQL, MariaDB, SQL Server, and SQLite through one interface. Add it when you want one server for read-mostly querying across several databases instead of a different one per engine.
• MCP Toolbox for Databases — Google's open-source googleapis/genai-toolbox, connecting to self-managed Postgres/MySQL/SQLite plus managed engines (Cloud SQL, AlloyDB, Spanner, BigQuery, Neo4j, and more). Add it when your data lives across cloud databases and you want one governed layer in front of them.
• Filesystem (from the reference set above) doubles as your "data" server for anything that's just files on disk — CSVs, JSON, logs. Add it when the "database" is really a folder.

Web & Browser

• Playwright MCP — Microsoft's microsoft/playwright-mcp, Apache-2.0. Drives Chromium/Firefox/WebKit and feeds the agent the page's accessibility tree instead of screenshots, so actions are cheap and deterministic. Add it when you need repeatable browser steps and assertions ("log in, click, verify text").
• Chrome DevTools MCP — Google's ChromeDevTools/chrome-devtools-mcp, Apache-2.0. Gives the agent the DevTools surface: console messages, network requests, performance traces, screenshots. Add it when you're debugging a live page ("why is this slow / what's erroring") rather than driving a scripted flow.
• Fetch (reference set) is the low-cost option when you only need the text of a page, no rendering. Add it when a full browser is overkill.
• Puppeteer MCP — Anthropic's original browser reference server, now in servers-archived. Chromium-only, screenshot-driven, small tool surface. Add it when you specifically want the minimal legacy behavior — otherwise prefer Playwright or Chrome DevTools above.

Productivity & APIs

This is the category where "open source" gets slippery, so read licenses before you assume. Many SaaS vendors now publish their own MCP servers — Notion, Sentry, Slack, Supabase, MongoDB and others. A good number keep the server code public on GitHub even though the service behind it is a paid product. The pattern to expect: the connector is free and often open-source, but it talks to an account you may have to pay for.

• GitHub's server (listed under Dev & Git) is the standout genuinely-open, genuinely-free option here — Actions, releases, and project boards all count as "productivity."
• For the vendor connectors, my rule is simple: check the repo's LICENSE file and confirm it's the official org, not a look-alike fork, before you trust it with a token. Directories are full of unofficial re-uploads.

Memory & Utility

• Memory (reference) — the knowledge-graph store again, and still the easiest free way to give an agent durable recall. Add it when you want persistence without standing up a vector database.
• Sequential Thinking (reference) — structured reasoning as a tool. Add it when answer quality on hard multi-step tasks matters more than latency.
• Time (reference) — small, boring, and fixes a real recurring bug. Add it when anything time-sensitive touches your agent.

How to Use This List

Resist the urge to connect all of them. Every MCP server you attach loads its tool descriptions into the system prompt on every request — a dozen servers can quietly burn thousands of tokens and raise the odds the model picks the wrong tool. Pick the two or three that map to what you actually do this month, then add more only when a task demands it. If you want the deeper trade-off math and the shortlist I'd install first, the best MCP servers of 2026 post covers it, and how to add an MCP server to Claude Code walks the config.

And if you find yourself reaching for an MCP server just to teach the agent a workflow — no live system, no state — that's usually a job for a skill instead. We keep a growing shelf of those at orangebot.ai/skills.

MCP is moving fast enough that any static list goes stale. We track the servers, skills, and releases that actually matter each week — the honest, no-hype version — in the OrangeBot weekly digest. Subscribe there and skip the 3,000-entry directories.