Ranked roundup · AI agents
Best MCP Server for AI Agents
There is no single MCP server for AI agents — the useful unit is a small stack, and most of it now ships from official vendors. Anthropic's Model Context Protocol project maintains seven reference servers — filesystem, memory, sequential thinking, fetch, git, time, and a test server — in one repo with 90,000 GitHub stars. Microsoft ships Playwright MCP for browser automation, and GitHub ships its own server for repo work. The community fills the gap that matters most for long-running agents, persistent memory, where Mem0 is the largest project — with the caveat that its documented MCP endpoint is now a hosted service, not the local OpenMemory stack it launched with. If you would rather run nothing at all, Composio bundles 1,400+ tool integrations behind one managed platform. We ranked the five below on verifiable criteria only: official status, commit recency, adoption, auth model, and coverage. Nobody paid for placement, and the unflattering parts — archived repos, sunset products — are reported as found.
Facts checked September 1, 2026 · ranked on verifiable criteria · placement is never for sale
Founder, Automation Squad ·
The ranking
- Official · Model Context Protocol project (Anthropic)
The starter stack — first three servers almost every local agent loads
Seven maintained servers in one repo: filesystem (scoped read/write), memory (local knowledge-graph persistence), sequential thinking (structured multi-step reasoning), fetch (web content to markdown), git, time, and a protocol test server ('everything'). Filesystem plus memory plus sequential thinking is the standard foundation for a local agent stack.
Local, stdio — each server installs separately via npx or uvx · None — no accounts, no API keys; servers run with your local user's permissions · MIT/Apache-2.0 (relicensing in progress per repo LICENSE) · 90,000★
- These are reference implementations meant to demonstrate the protocol, not hardened products
- No auth layer at all — the filesystem server's only guardrail is the directory allowlist you pass it, so scope it carefully
- The memory server persists to a local JSON knowledge graph — it does not sync across machines
- The repo has been trimmed over time to these seven; older bundled servers were moved out, so tutorials referencing others in this repo are stale
- License is mid-transition from MIT to Apache-2.0, so GitHub reports it as NOASSERTION
- Official · Microsoft
Agents that need to operate real websites — research, form-filling, end-to-end testing
Full browser automation for agents: navigation, clicking, typing, form filling, screenshots, and structured page snapshots via the accessibility tree rather than pixel-matching, which keeps token usage down and makes actions deterministic.
Local — one npx command; runs a real browser on your machine · None for the server itself; sites your agent logs into use your local browser profile/session · Apache-2.0 · 36,697★
- Runs a real browser locally, which is heavier than API-based tools
- Browser automation only — it is one layer of an agent stack, not a general toolkit
- No hosted option from Microsoft; you run and update it yourself
- Pushed daily at time of checking — active is good, but pin versions if you need stability
- Official · GitHub
Agents that read, write, and review code as part of their loop
Repos, issues, pull requests, code search, Actions — the operational surface an agent needs to actually ship work. The hosted option means no install and no token management.
Hosted remote server from GitHub, or a local binary/Docker image you run yourself · OAuth for the hosted server; a personal access token you scope yourself for local · MIT · 32,645★
- Domain-specific — it earns a slot in an agent stack only if the agent touches GitHub
- A carelessly scoped PAT gives the agent broad write access to your repos; use fine-grained tokens
- The hosted server sends your agent's repo operations through GitHub's remote endpoint, which some orgs restrict
- Community
Agents that need memory shared across sessions, machines, and multiple MCP clients
Persistent cross-session memory for agents: 11 tools covering add, semantic search, filtered retrieval, update, single and bulk delete, entity management, and event tracking. Memory survives across sessions, clients, and machines because it lives in your Mem0 account.
Hosted only — documented endpoint is https://mcp.mem0.ai/mcp; nothing runs on your machine · Mem0 Platform account required; browser OAuth sign-in or an API key as bearer token for headless use · Apache-2.0 · 64,511★
- The standalone mem0-mcp repo was archived in March 2026
- OpenMemory, the local Docker-based memory stack Mem0 launched with, was removed from the monorepo in July 2026 — guides describing a local-first OpenMemory MCP are out of date
- The documented MCP path is hosted-only: memories live in Mem0's cloud, not on your machine
- Requires a Mem0 Platform account; the 64,511 stars belong to the Mem0 memory library monorepo, not a standalone MCP server repo
- For fully local memory with no account, the official reference memory server (rank 1) is the alternative, at the cost of no cross-machine sync
- Community
Wiring many SaaS tools into an agent without managing servers or OAuth flows
One managed layer over roughly 1,400 app toolkits (Gmail, Slack, Notion, CRMs, and so on) exposed to agents via MCP or direct APIs, with tool search and context management on top. The pitch is breadth without running or authenticating anything yourself.
Managed cloud platform — Composio hosts the servers and the integration layer · Composio account; the platform manages authentication to third-party apps on your behalf · MIT · 29,995★
- Commercial hosted platform — a free tier exists, but it is a product, not infrastructure you own
- Your third-party OAuth grants and tokens flow through Composio's platform, which is a real trust decision
- Marketing says 20,000+ tools while the toolkit directory lists 1,442 integrations — the counts measure different things, so verify your specific app is covered
- Klavis AI is the nearest alternative in this category but was excluded: last push 2026-06-01, three months stale at verification
AI agents MCP servers compared
| Server | Maintained by | Runs | Auth | License |
|---|---|---|---|---|
| MCP Reference Servers (Filesystem, Memory, Sequential Thinking) | Model Context Protocol project (Anthropic) | Local, stdio — each server installs separately via npx or uvx | None — no accounts, no API keys; servers run with your local user's permissions | MIT/Apache-2.0 (relicensing in progress per repo LICENSE) |
| Playwright MCP | Microsoft | Local — one npx command; runs a real browser on your machine | None for the server itself; sites your agent logs into use your local browser profile/session | Apache-2.0 |
| GitHub MCP Server | GitHub | Hosted remote server from GitHub, or a local binary/Docker image you run yourself | OAuth for the hosted server; a personal access token you scope yourself for local | MIT |
| Mem0 MCP | Community | Hosted only — documented endpoint is https://mcp.mem0.ai/mcp; nothing runs on your machine | Mem0 Platform account required; browser OAuth sign-in or an API key as bearer token for headless use | Apache-2.0 |
| Composio | Community | Managed cloud platform — Composio hosts the servers and the integration layer | Composio account; the platform manages authentication to third-party apps on your behalf | MIT |
How to choose
Pick the reference servers if you are building your first agent stack — filesystem, memory, and sequential thinking cover most of what a local agent needs, with no accounts and no keys. Add Playwright MCP if your agent has to drive a browser; it runs locally from one npx command. Pick GitHub's server if the agent works in repos — the hosted version handles OAuth for you. Pick Mem0 if you need memory that survives across sessions and machines, and accept that this now means a hosted platform account. Pick Composio if you want many SaaS integrations without managing credentials yourself.
How this ranking works
Every server here is real and was checked against its repository or vendor documentation on the date above — stars and licenses come from the GitHub API, not from memory. Order reflects verifiable criteria only: official support, maintenance recency, adoption, auth model, and documented capability coverage. We do not run paid placements, and we list every caveat we find, including the unflattering ones.
Common questions
- Is there an official MCP server for AI agents?
- Several, depending on the layer. Anthropic's Model Context Protocol project maintains seven reference servers (filesystem, memory, and sequential thinking among them), Microsoft maintains Playwright MCP for browser automation, and GitHub maintains its own server — all three repos had commits within 24 hours of our 2026-09-01 check.
- What does setup and auth actually look like for an agent stack?
- The reference servers and Playwright MCP run locally over stdio from a single npx or uvx command with no accounts or API keys. GitHub's hosted server uses OAuth (or a personal access token if you run it locally), Mem0's endpoint requires a Mem0 Platform account with browser sign-in or an API key, and Composio manages third-party OAuth for you on its hosted platform.
- What happened to OpenMemory?
- Mem0 archived the standalone mem0-mcp repo in March 2026 and removed the local Docker-based OpenMemory stack from its monorepo in July 2026. The documented path today is Mem0's hosted MCP server at mcp.mem0.ai, which stores memories in your Mem0 account rather than on your machine.
- Do I need all seven reference servers?
- No — each installs as a separate package from the same repo. Filesystem, memory, and sequential thinking are the usual starting three; fetch, git, and time are worth adding only when your agent actually needs them, and 'everything' is a protocol test server, not for production.
MCP servers we build
Automation Squad publishes its own open-source MCP servers and agent skills, documented with the same rules this ranking uses.