Best Open-Source MCP Servers in 2026: 10 Projects Compared

The best open-source MCP servers are maintained by the companies or projects behind the tools they connect: the official MCP reference servers, GitHub, Microsoft's Playwright, Chrome DevTools, Grafana, HashiCorp Terraform, MongoDB, the Kubernetes server, Firecrawl and AWS Labs. All ten publish their code under a standard license, so you can read what the server does before you run it.

Open source matters more for MCP than for most software. An MCP server sits between an AI and your systems, and its code decides what the AI can touch. With public code, your developers can check for hidden data collection, pin a version and run it inside your own network.

We checked each project against its official repository on 28 October 2026. eSEOspace did not install or test these servers for this list.

How we chose

We included a server only if it passed these tests:

1. Open license. The repository states an OSI-style license such as MIT, Apache-2.0, MPL-2.0 or GPL.
2. Maintained by an accountable owner. A company or the MCP project itself maintains it, or it is an established community project.
3. Live and documented. We could open its README on 28 October 2026 and it was not archived.
4. Documented security basics. The README explains sign-in, permissions or known risks.

We skipped servers that exist only as copies in directories, and projects the maintainers have archived. For vendor-hosted options instead, see our list of remote MCP servers. For a developer-focused list, see MCP servers for developers.

Comparison table

1. MCP reference servers: the official teaching set

What it is. The modelcontextprotocol/servers repository holds seven reference servers: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. The repo says they are "meant to serve as educational examples for developers building their own MCP servers, not as production-ready solutions." Older servers such as GitHub, Slack and PostgreSQL were moved to a separate archive repo.

Best for: Developers learning MCP, and simple local jobs such as reading files in one folder.

Pricing: Free. The project uses Apache-2.0 for new contributions, with existing code under MIT.

Pros:

• Maintained alongside the MCP specification itself.
• Small, readable code that shows how servers should behave.
• Filesystem access can be limited to chosen folders.

Cons:

• Explicitly not production-ready.
• Many older servers in tutorials are now archived, which causes confusion.

Checked against: MCP servers repository on 28 October 2026. Profiles: Filesystem and Fetch.

2. GitHub MCP Server: open code for GitHub's own server

What it is. The GitHub MCP Server is GitHub's official server, published under the MIT license. You can run it locally with a personal access token or use the hosted endpoint. Toolsets control which groups of tools load, and --read-only skips write tools.

Best for: Teams that want GitHub access with code they can audit.

Pricing: Free and open source. GitHub's docs say it is "available to all GitHub users regardless of plan type."

Pros:

• Same code runs locally or as GitHub's remote server.
• Read-only mode takes priority over requested write tools.
• Toolsets keep the tool list focused.

Cons:

• Personal access tokens in local setups are long-lived secrets.
• Some tools need paid GitHub or Copilot features.

Checked against: GitHub MCP Server README on 28 October 2026. Full profile: GitHub MCP server.

3. Playwright MCP: browser automation from Microsoft

What it is. Playwright MCP "enables LLMs to interact with web pages through structured accessibility snapshots." It uses Playwright's accessibility tree instead of screenshots, so it needs no vision model. You run it with npx @playwright/mcp@latest, and it can also serve over HTTP on a local port.

Best for: QA and web teams that want an AI to click through and test pages.

Pricing: Free and open source under Apache-2.0.

Pros:

• Fast, text-based page snapshots.
• File access is limited to workspace roots by default.
• Backed by the Playwright team at Microsoft.

Cons:

• The README says it "is not a security boundary."
• The browser_run_code_unsafe tool is "RCE-equivalent," so keep it off.

Checked against: Playwright MCP on GitHub on 28 October 2026. Full profile: Playwright MCP server.

4. Chrome DevTools MCP: debugging a live Chrome browser

What it is. Chrome DevTools MCP lets a coding agent "control and inspect a live Chrome browser." It covers performance traces, network and console debugging, screenshots and automation built on Puppeteer. It runs with npx -y chrome-devtools-mcp@latest.

Best for: Developers fixing page speed and front-end bugs.

Pricing: Free and open source under Apache-2.0.

Pros:

• Performance tracing that few other servers offer.
• Opt-out flags for usage statistics, CrUX lookups and update checks.
• Maintained by the Chrome DevTools team.

Cons:

• Usage statistics collection is on by default.
• It "exposes content of the browser instance to the MCP clients," including anything you are logged into.

Checked against: Chrome DevTools MCP on GitHub on 28 October 2026. Full profile: Chrome DevTools MCP server.

5. Grafana MCP: dashboards, metrics and logs

What it is. mcp-grafana connects an AI to Grafana dashboards, Prometheus, Loki and Grafana Incident. It signs in with a Grafana service account token. You can run it with uvx, Docker, a binary or from source.

Best for: On-call and platform teams that investigate incidents in Grafana.

Pricing: Free and open source under Apache-2.0.

Pros:

• --disable-write blocks every write tool and removes raw SQL tools.
• Service account tokens let you scope access in Grafana.
• Covers queries, dashboards and incidents in one server.

Cons:

• Needs Grafana 9.0 or later for full features.
• Write mode can change dashboards and incidents, so default to read-only.

Checked against: grafana/mcp-grafana README on 28 October 2026. Full profile: Grafana MCP server.

6. Terraform MCP Server: infrastructure as code from HashiCorp

What it is. The Terraform MCP Server connects an AI to the Terraform Registry and to HCP Terraform or Terraform Enterprise. It can look up providers and modules and manage workspaces, variables and runs. It runs in Docker or from Go, and supports stdio and Streamable HTTP.

Best for: DevOps teams that write and review Terraform.

Pricing: Free and open source under MPL-2.0. A hosted endpoint at https://mcp.terraform.io/mcp is also listed.

Pros:

• In HTTP mode, each user passes their own token, so RBAC follows the person.
• An organization allowlist limits which HCP Terraform orgs it can reach.
• Optional OpenTelemetry metrics for tool calls.

Cons:

• The README warns against using it with untrusted clients or models.
• Workspace delete tools exist, so token scope matters.

Checked against: Terraform MCP Server README on 28 October 2026. Full profile: Terraform MCP server.

7. MongoDB MCP Server: databases and Atlas

What it is. The MongoDB MCP Server is "a Model Context Protocol server to connect to MongoDB databases and MongoDB Atlas Clusters." It connects with a connection string, or an Atlas service account for Atlas tools. MongoDB also runs an Atlas-hosted remote version that uses OAuth.

Best for: Developers who query and manage MongoDB data.

Pricing: Free and open source under Apache-2.0.

Pros:

• --readOnly allows only read, connect and metadata operations.
• Docs recommend environment variables over command-line secrets.
• Local and hosted options from the same vendor.

Cons:

• Telemetry is on by default (turn it off with MDB_MCP_TELEMETRY=disabled or DO_NOT_TRACK=1).
• Write mode can change production data if you point it at the wrong cluster.

Checked against: MongoDB MCP Server README on 28 October 2026. Full profile: MongoDB MCP server.

8. Kubernetes MCP Server: native cluster access

What it is. The Kubernetes MCP Server from the containers project is "a Go-based native implementation that interacts directly with the Kubernetes API server." It does not wrap kubectl or helm, so nothing else needs installing. It ships as npm, Python, binary and container packages.

Best for: Platform engineers who want an AI to inspect clusters.

Pricing: Free and open source under Apache-2.0.

Pros:

• read_only = true and denied_resources settings, for example to hide Secrets.
• OAuth/OIDC sign-in for HTTP mode, with Keycloak and Entra ID guides.
• Many install options.

Cons:

• Community-maintained rather than published by the Kubernetes project itself.
• Uses your kubeconfig by default, which may carry admin rights.

Checked against: kubernetes-mcp-server README on 28 October 2026. Full profile: Kubernetes MCP server.

9. Firecrawl MCP Server: web scraping and search

What it is. The Firecrawl MCP Server lets clients "search, scrape, and interact with the live web" and get Markdown or structured data back. It runs locally, can point at a self-hosted Firecrawl API, or connects to the hosted endpoint at https://mcp.firecrawl.dev/v2/mcp.

Best for: Research and content teams that need clean web data in an AI chat.

Pricing: The server is MIT-licensed. Calls to Firecrawl's API use credits: Firecrawl's pricing page lists Free (1,000 credits a month), Hobby ($19/month), Standard ($99/month), Growth ($399/month), Scale ($749/month) and custom Enterprise.

Pros:

• Keyless free tier for scrape, search and parse, with rate limits.
• Self-hosting option keeps scraping in your own environment.
• Clear warning never to put API keys in the server URL.

Cons:

• The open-source server mostly calls a paid API.
• Scraped pages can carry prompt-injection text into your AI.

Checked against: Firecrawl MCP Server README and Firecrawl pricing on 28 October 2026. Full profile: Firecrawl MCP server.

10. AWS Labs MCP servers: a large collection for AWS

What it is. The awslabs/mcp repository holds many open-source MCP servers for AWS, grouped by use case: documentation, infrastructure, containers, serverless, databases, cost and more. The README says "The Agent Toolkit for AWS is the successor" to these servers, and that the most useful projects will move there over time. It adds that "This repo continues to work and accept contributions."

Best for: AWS teams that want a specific, narrow server they can run themselves.

Pricing: Free and open source under Apache-2.0. You still pay normal AWS charges for the resources the AI uses.

Pros:

• Narrow servers let you grant only the AWS access a job needs.
• Maintained in AWS's own GitHub organization.
• Open code alongside AWS's hosted server.

Cons:

• Some servers are deprecated, such as the Cloud Control API server.
• The move to Agent Toolkit for AWS means some servers may change home.

Checked against: awslabs/mcp on GitHub on 28 October 2026. Full profile: AWS MCP server.

What to check before you run an open-source server

• Who maintains it. Prefer the vendor's own organization or the MCP project. Copies with similar names are a known supply chain risk.
• Pin a version. @latest in a config file means new code runs without review.
• Start read-only. Use the read-only flags listed above, then widen access later.
• Check telemetry. Read the README for usage statistics and opt-out flags.
• Run it in a sandbox. Our page on sandboxing local servers explains how.

More picks by category are in the best-of collection.

Next step

If you need a server that does not exist yet, or a hardened fork of an open-source one, our MCP development team builds and reviews servers for business use.

Frequently asked questions