12 Best MCP Servers for Automation in 2026: Official Servers, Registries and Automation Platforms with MCP Compared

Quick answer: the best MCP servers for automation by need

For automation work in 2026, start with Zapier MCP to reach thousands of apps through one automation platform, or Composio for a managed tool catalogue. GitHub covers code and repositories, Playwright browser automation, Supabase databases, Stripe payments, Notion docs. Atlassian handles Jira and Confluence, HubSpot CRM, Cloudflare edge infrastructure. The official reference servers cover files, memory and everyday tasks as examples rather than production tools, and IFTTT covers consumer automations.

Every server below comes from the vendor's own documentation or the official protocol repositories and registry, read in September 2026. Nothing here draws on a third-party catalogue, and there are no ratings, popularity counts or tested claims. The numbering is an order of presentation, not a ranking.

What an MCP server is and what makes one good for automation

An MCP server is an application that exposes tools, prompts and resources to an MCP client. That is Cloudflare's definition. The client is your AI tool: an editor, a CLI, an agent harness. The server reads a repository, runs a query or files a ticket.

Servers come in two shapes, as Cloudflare's docs put it: remote connections over the Internet using Streamable HTTP, and local ones on the same machine using stdio. Remote servers usually hand authentication to the vendor through OAuth. Local ones read a token from your environment.

For automation, four things matter more than the length of a tool list. Does the server write as well as read? Does the vendor handle auth? Can writes be approved before they run? And how is usage metered? Some servers cost nothing beyond the product plan. Others count every call against an allowance, so each entry below names what the page states.

MCP servers at a glance

1. Composio: best managed tool platform with MCP servers for many apps

Composio sells a managed tool catalogue that an agent reaches through one endpoint. As of September 2026 its pricing page puts the Free plan at $0 a month, with 100,000 tool calls, 50,000 trigger events and 3 team members. The page says that pricing applies to signups on or after 15 August 2026, with earlier customers grandfathered through 31 December 2026. Free is hard-capped: at a cap, usage pauses until the next month or an upgrade, at no charge.

Scale is $29 a month and includes a $29 monthly usage credit: unused credit expires at the end of each billing month and does not roll over. Each tool execution counts once, and meta tools such as tool search are free. Past the 100,000 free calls, Scale bills $0.0003 per tool call, and tools on Composio's shared managed apps add $0.0002 a call after 20,000 free calls a month. Composio markets its MCP Gateway as one managed endpoint for a customer's tools and agents. That is the vendor's description, not a tested capability.

2. Zapier MCP: best for reaching thousands of apps from one server

Zapier pitches its MCP server as one connection that gives an AI tool access to more than 9,000 apps. The quickstart gives Claude, Claude Code, ChatGPT and Cursor their own connection pages. It lists twenty more supported clients, among them Gemini CLI, Microsoft Copilot Studio, VS Code, Warp, Windsurf and Zed. You need a Zapier account and a client that speaks the protocol over Streamable HTTP.

"Zaps run automatically based on triggers — when an email arrives, when a form is submitted. MCP is different: it lets your AI take actions on demand, based on what you ask it to do in the moment. Think of Zaps as scheduled automation and MCP as giving your AI hands. Most people use both."

Source: zapier.com

Billing catches people out. There is no separate product, so MCP runs on your existing plan. Each successful tool call spends two tasks from the same monthly allowance your Zaps use, and failed calls consume none. Batch work multiplies, as Zapier's examples show: five spreadsheet rows are five tool calls and ten tasks. The task allowance is the only usage limit, and tasks reset monthly.

The pages disagree on one point. zapier.com/mcp badges the product Beta in its hero, while zapier.com/pricing lists it as a feature of all four plans, Free, Professional, Team and Enterprise, with no beta marker.

3. GitHub MCP server: best for code and repositories

GitHub's own server connects AI tools to GitHub so an agent can read repositories and code files, manage issues and pull requests, analyse code and automate workflows. GitHub hosts a remote version and offers a local one for hosts without remote support. On github.com it logs you in with OAuth on first use. A personal access token set in GITHUB_PERSONAL_ACCESS_TOKEN takes precedence over OAuth.

Tools are grouped into toolsets you switch on or off with a toolsets flag. Specify nothing and the server loads context, repos, issues, pull_requests and users. A read-only flag offers read tools only and prevents modifications to repositories, issues and pull requests. Lockdown mode filters content surfaced from public repositories to reduce prompt-injection risk. Still, the README says it is not an authorization boundary. The repository advises granting only the minimum token scopes needed. It names no plan or price, only that any applicable policies must be enabled.

4. Playwright MCP: best for browser automation

Playwright MCP gives browser automation capabilities using Playwright. The defining choice is how a page is represented: LLMs interact with web pages through structured accessibility snapshots rather than screenshots. It works from the accessibility tree instead of pixel-based input. No vision model is needed, and the server operates purely on structured data.

It requires Node.js 18 or newer and an MCP client. The README names VS Code, Cursor, Windsurf, Claude Desktop, Goose, Grok and Junie. By default it runs with a persistent browser profile. Isolated sessions are an option for testing, or the server can connect to an existing browser through the Playwright MCP browser extension. The README names no browser brand for it. It states no cost for the server.

5. Supabase MCP server: best for databases and backends

Supabase opens its MCP guide with a warning rather than a feature list. Connecting an LLM to your Supabase projects carries security risks, and you should read the security best practices first. The database tools include execute_sql and apply_migration alongside list_tables, list_extensions and list_migrations.

Tools sit in feature groups, and every group except Storage is enabled by default. The hosted server authenticates in a browser window: you sign in and grant the MCP client access to the organization, with no personal access token. Two query parameters matter. Set read_only to true and all queries run as a read-only Postgres user. Set project_ref and the server is scoped to one project, with account tools disabled. Supabase notes that most clients ask you to accept each tool call before it runs, and advises keeping manual approval on. As of September 2026 the experimental branching group requires a paid plan. The page names no tier, price or minimum, so check Supabase's pricing page and confirm before purchase.

6. Stripe MCP: best for payments

Stripe documents its MCP server as being in public preview as of September 2026, so treat the behaviour below as subject to change. It lets AI agents interact with the Stripe API and search Stripe's knowledge base, including documentation and support articles. The published tools include stripe_api_read for any Stripe API GET method, stripe_api_write for POST, PATCH, PUT and DELETE methods, and search_stripe_documentation. That list is the preview one and can change.

It is a remote server at https://mcp.stripe.com that uses OAuth to authorize clients, so access can be granted and revoked without sharing an API key, and clients without OAuth can send a restricted API key as a bearer token. It is a preview, so check the address in Stripe's docs first. In practice, two rules matter. Certain write actions such as refunds and outbound payments require human confirmation. An unapproved action expires after 24 hours. Both are preview behaviour. The page states no pricing, quota or rate limit.

7. Notion MCP: best for docs and wikis

Notion hosts its MCP server remotely, so there is nothing to install. Clients such as Claude Code, Cursor or Codex connect to a workspace once the user authorizes the connection with OAuth. The tools search content in Notion and connected sources, read, create and update Notion content, and create pages and databases. That covers most of what a documentation agent needs.

Workspace owners manage MCP client access in Settings, then Connections. Organization owners can use the Admin API to list and revoke members' connections, which lets you audit who wired an agent into the workspace. Notion's documentation names no plan requirement and no price. If a specific tier matters, confirm it with Notion.

8. Atlassian MCP server: best for Jira and Confluence

The Atlassian server lets an agent summarise and search Jira, Jira Service Management, Confluence, Bitbucket, Projects and Goals. It can also create or update work items and pages. It is remote, reached at https://mcp.atlassian.com/v2/mcp, and supports any app with MCP support: OpenAI ChatGPT, Claude, Docker, GitHub Copilot CLI, Google Gemini and Amazon Quick Suite.

Authorization uses OAuth 2.1, so actions respect the user's existing access controls and permissions. API token authentication is also available. Atlassian warns that MCP clients can act on all connected products with those permissions, and advises least privilege and monitoring audit logs.

As of September 2026, according to Atlassian's support documentation, version 2 has been released, and from 1 March 2027 existing v1 use will automatically expose v2 tools. That date comes from an undated support page, so confirm it with Atlassian. According to the same documentation, calls that retrieve data or generate insights consume Rovo credits from the organization's shared pool. The page states no thresholds, allowances or reset behaviour.

9. HubSpot MCP server: best for CRM data

HubSpot publishes two servers: a remote one for a HubSpot CRM account, and a local CLI-based Developer MCP server for the HubSpot Developer Platform. The remote server requires the new HubSpot Developer Platform and a user-level application with read scopes for the CRM objects you need. OAuth credentials connect it to the mcp.hubspot.com endpoint.

HubSpot's page contradicts itself on writes. The CRM-data FAQ answer lists read and write access for contacts, companies, deals, tickets, orders, invoices, quotes and subscriptions, plus engagements such as calls, emails, meetings, notes and tasks. The sensitive-data FAQ answer on the same page says the server has read-only access to standard records such as contacts, companies, deals and tickets. It also says there is no access to custom Sensitive Data Properties. Confirm with HubSpot before relying on write access. The page carries a forward-looking line about aligning with OAuth 2.1, so only OAuth 2.0 can be confirmed, and it names no price or seat requirement of its own.

10. Cloudflare MCP servers: best for edge and infrastructure

Cloudflare's Agents documentation is written for people building and deploying MCP servers on Cloudflare. It also carries the clearest short definition: MCP servers are applications that expose tools, prompts and resources to MCP clients. The same page draws the transport line. Remote connections run over the Internet using Streamable HTTP, local ones on the same machine using stdio.

Cloudflare gives two pieces of design advice. It recommends deploying several focused MCP servers with narrowly scoped permissions, which reduces over-privileged access and makes auditing easier. It also advises against treating a server as a wrapper around a full API schema, and recommends tools built for specific user goals. The page links to Cloudflare's own MCP servers rather than listing them. That section describes a catalogue of managed remote servers reached with OAuth: documentation, Workers bindings, observability, Radar and audit logs. Neither page states a cost.

11. Reference servers: best for files, memory and everyday tasks

The modelcontextprotocol/servers repository collects reference implementations of the protocol plus references to community-built servers and additional resources. It is managed by Anthropic and built together with the community.

Seven reference servers are current: Everything (a reference and test server with prompts, resources and tools), Fetch (web content fetching and conversion) and Filesystem (secure file operations). The rest are Git (reading, searching and manipulating Git repositories), Memory (knowledge graph-based persistent memory), Sequential Thinking (problem-solving through thought sequences) and Time (time and timezone conversion). The modelcontextprotocol.io examples page lists the same seven as official reference servers demonstrating core MCP features and SDK usage.

These are educational examples for developers building their own servers, not production-ready solutions. According to the official MCP servers repository, an archived section holds earlier reference servers, among them GitHub, GitLab, Google Drive, Google Maps, PostgreSQL, Puppeteer, Redis, Sentry, Slack and SQLite, all moved to a servers-archived repository. The repository points readers to the MCP Registry for published servers, and keeps only the small number of reference servers maintained by the MCP steering group.

12. IFTTT MCP: best for consumer automations

IFTTT's MCP page describes a server that lets AI tools search, build and run automations across more than 1,000 services and more than 50 million applets. That is a different job from the single-app servers above. The agent assembles consumer automations rather than calling one API. The page's call to action is to sign up for free to IFTTT.

The setup steps mention a paid plan. According to IFTTT's MCP page, that is the AI tool's own plan rather than IFTTT's: the Claude instructions tell the reader to be signed up for a paid Claude plan. The page names no IFTTT plan at all, so no IFTTT tier can be described as required or not required. It does not state whether the server runs remotely or locally.

Where Latenode fits: MCP for 800+ apps and for building workflows

Latenode, which publishes this blog, is not one of the twelve above. As of September 2026 its MCP page connects agents such as Cursor, Claude Desktop, Claude Code, VS Code and Codex to 800+ apps through one server endpoint with managed OAuth: when the agent needs a service it sends you a link, you click Allow, and Latenode stores and refreshes the credentials. It is free to start. Premium apps, 23 apps with 69 actions on Latenode's own contracts, are billed pay as you go from Latenode credits, and 324 data providers sit behind the same server on one pay-as-you-go bill.

A second server, the Latenode MCP Server, connects an agent to a Latenode workspace so it can create, update and activate scenarios, search node types and connections, and run a scenario or a single node; Latenode recommends pairing it with its latenode-scenario-builder Skill. Neither page publishes a per-call price, so confirm the pay-as-you-go rates before budgeting.

How to choose and connect an MCP server safely

Start with servers the vendor maintains, then read what the tools can do: a server that runs SQL can drop a table. Copy the Supabase pattern. Scope the server to one project, run queries as a read-only user, and keep manual approval on, since most clients ask you to accept each tool call.

Use least-privilege tokens. GitHub advises granting only the minimum scopes needed, Atlassian warns that a client acts with the user's existing permissions, and Cloudflare prefers several narrowly scoped servers to one broad one. Know the meter before scaling a loop: Zapier spends two tasks per successful call, Composio counts each tool execution, and Atlassian consumes Rovo credits with no published threshold. Prefer remote when the vendor should hold auth, local when data must stay on the machine.

For discovery, use the official registry at registry.modelcontextprotocol.io, built in the open by MCP contributors in its own words. As of September 2026 its home page listed recently updated servers with versions and dates. Smithery, as of September 2026, describes itself as a place to connect agents to thousands of tools with auth, credentials and sessions handled for the user, and says it is now part of Arcade.dev. That is a self-description, nothing more. Ranked catalogues are not sources here. The consent rules above come from vendor pages, not the specification, which showed version 2026-07-28 as the latest in September 2026.

Which servers work together and which overlap

Single-app servers stack cleanly. GitHub, Supabase, Notion and Stripe each own a different system, and each carries its own auth. One client can hold several at once without them colliding. Scope each narrowly and the combination is safer than one broad server, which is the pattern Cloudflare recommends.

The overlap sits between the platform servers. Zapier MCP and Composio chase the same job, one endpoint reaching many apps, so running both usually means paying twice for the same reach. They also overlap with single-app servers: a platform connector can act in Gmail or Slack, while a dedicated server for a product almost always exposes deeper tools for it.

Use a platform server for a long tail of shallow actions across many apps, a dedicated server when the work goes deep into one product. When a job needs multi-step logic, branching or retries, that belongs on a workflow platform rather than inside one agent tool call.