Best list · Updated August 2026
Best MCP Servers for AI Agents
Best MCP servers for AI agents: the servers OpenSourcesAI has installed, discovered and executed itself, with security scope and install entry.
Editorial review
AI tools, model releases, pricing, licenses, and platform terms can change quickly. Verify the official source before production or commercial use.
Who this page is for
This page lists only MCP servers we have run ourselves. Every pick below is a Hands-on Verified profile in the MCP servers directory: at least one recorded configuration where installation, discovery and execution were verified in a named environment. Servers we have only read the source of are profiled there as Source Validated and are deliberately not picks here — the list grows as verifications land, and it never contains a server whose upstream package is deprecated, archived or unpublished at build time.
Selection criteria
- Hands-on Verified: at least one recorded configuration where installation, discovery and execution were verified in a named environment — never installed-only, never inferred.
- Install path still live upstream at build time (not deprecated, archived or unpublished) — a server that goes stale drops off this list automatically.
- Security scope stated up front — read-only, read-write or local execution — with an isolated or read-only way to start.
- Solves a narrow agent workflow (live docs, a browser, a debugger) rather than promising everything.
- Works with common MCP-capable clients — Claude Code, Claude Desktop, Cursor, Cline — with a config entry you can paste.
Top picks
Only servers OpenSourcesAI has run itself appear here — each pick is a Hands-on Verified profile with its configuration, security scope and config entry recorded. The rest of the profiles, including the Source Validated ones, are in the MCP servers directory.
1. Hands-on Verified · Read-only · Documentation + context
Context7
Upstash's MCP server that puts current, version-specific library docs and code examples in the prompt so a coding agent stops guessing at APIs.
Best for
Anyone whose coding agent keeps suggesting removed methods or outdated signatures for fast-moving libraries — Next.js, React, LangChain — and wants live docs in the prompt without building a RAG pipeline.
Watch out
Every query goes to Upstash's hosted API — the local package is a thin shim and the backend is not open source, so it is not for air-gapped work. Coverage depends on the library being indexed; internal or private libraries need an API key and a private repo. It grounds documentation only: it does not run, lint or test code.
2. Hands-on Verified · Local execution · Browser + DevTools
Playwright MCP
Microsoft's MCP server that drives a real browser through Playwright with accessibility snapshots, not screenshots — navigate, click, type, read.
Best for
Agents that need to actually use a website — fill forms, click through flows, read rendered pages — with a browser you can watch, on your own machine.
Watch out
It executes actions in a real browser on your machine and one tool (browser_run_code_unsafe) is documented as RCE-equivalent; the README states outright that Playwright MCP is not a security boundary and that origin allow/block lists do not affect redirects. Use --isolated for a throwaway profile, keep approval prompts on, and do not point it at a browser profile that is signed into anything you care about. Not for production automation.
3. Hands-on Verified · Local execution · Browser + DevTools
Chrome DevTools MCP
The Chrome DevTools team's MCP server: DOM snapshots, console, network and performance traces from a live Chrome as structured tool results.
Best for
Frontend and full-stack agent workflows where the agent should see the real DOM, console errors and network waterfall after a change instead of a screenshot loop.
Watch out
It drives a real browser on your machine. Attaching to a Chrome you already use means opening a debugging port that any local application can connect to — keep that to test profiles, never a signed-in daily browser. Chrome-only; no Firefox or Safari.
How to choose
Start with the read-only pick — a documentation server changes nothing on your machine and shows you the client's approval flow. Add a browser or DevTools server only in an isolated profile, keep approval prompts on, and never point it at a browser that is signed into anything you care about. The best MCP server is the one that solves a narrow workflow with clear review boundaries; the full set of profiles, including the ones we have not run yet, is in the MCP servers directory.
Related links
OpenSourcesAI may use affiliate links or sponsored placements in the future. Sponsored placements should be clearly labeled, and affiliate relationships should not guarantee positive coverage.
FAQ
Which MCP server should beginners try first?
Start with a read-only server. Of the picks here, a documentation server is the gentlest: it changes nothing on your machine and shows you how the client asks for approval. Add a browser or DevTools server afterwards, in an isolated profile.
Why are only a few servers listed?
Because a pick here means we ran it — installed, discovered and executed in a recorded environment. Servers we have only read the source of are profiled in the MCP servers directory as Source Validated, and they move onto this list when a verification is recorded, not before.
Do MCP servers replace normal app integrations?
No. MCP is useful for agent tool access, but stable production workflows may still need direct APIs, tests, permissions, and monitoring.
Related resources
Continue comparing tools, models, stacks, and guides related to this category.
Sources
Sponsorship note
Built an AI tool or open-source project? Submit it for review or sponsor a featured placement on OpenSourcesAI.
Sponsor or submit