The 21 Best MCP Servers in 2026 (Tested & Compared)

1. La Growth Machine, best for running outreach from an assistant

Overview. La Growth Machine is a multichannel outreach platform, and its MCP server is the connection between that platform and an assistant. We run it in our own workspace, which is where the tool count below comes from. It covers the contact and post-contact layer: building audiences, launching sequences, handling replies and reading performance.

Key features. 41 actions: building an audience from a Sales Navigator search, listing and analysing campaigns, pulling live stats, fetching conversations and reply history, drafting and sending LinkedIn or email replies, and managing sender identities. Stats come back broken down by channel and by sequence step, with deal stages including calls booked. Write actions require explicit confirmation by default, and a read-only mode is available.

Pros & cons. ✅ The widest set of outreach actions here, and it works with any MCP client: Claude, ChatGPT, Claude Code, Codex, Cursor, Windsurf, GitHub Copilot and n8n ✅ Open source and free, with write actions gated behind confirmation ⚠️ It needs an active account, so it is not something you try in isolation ⚠️ It covers execution and reply handling, not upstream list sourcing, which stays with AI prospecting tools

Pricing. The server is free and open source. The platform starts from €60/month per identity (Basic, monthly billing, verified August 2026), with a 14-day free trial and no card required. Enrichment runs on a monthly credit allowance, 1,250 credits on Basic, and the REST API starts at Pro.

Best for. B2B teams running LinkedIn and email sequences who want to build, launch and answer campaigns from their assistant. Details on the La Growth Machine MCP page.

2. GitHub, best for code, issues and pull requests

Overview. GitHub’s official server connects an assistant to repositories, and almost every host supports it. We did not run it ourselves, so what follows comes from GitHub’s documentation.

Key features. Reading repositories, searching code, opening and triaging issues, submitting pull requests. OAuth, with scopes set per token.

Pros & cons. ✅ Universally supported, so it works wherever your assistant runs ✅ Scoped OAuth tokens limit blast radius ⚠️ One of the heavier servers in tool count, a real cost when you run several ⚠️ Not measured in our stack, so no field verdict here

Pricing. Free with a GitHub account, verified August 2026.

Best for. Engineering teams who want code review, triage and repository questions handled in the same place they write.

3. Notion, best for team knowledge

Overview. Notion’s official server reads and writes pages, which makes a workspace queryable and editable by an assistant. We run it, and it exposes 28 tools as of August 2026.

Key features. Fetching and searching pages, creating and updating pages and databases, querying data sources, managing comments and users.

Pros & cons. ✅ Two-way: an assistant pulls context and writes results back, so notes stop being a dead end ✅ OAuth with workspace-level consent ⚠️ Query quotas apply on some data-source calls, easy to hit in a long session ⚠️ 28 tools is a lot of context if Notion is not central to your work

Pricing. Free with a Notion account; workspace plans start at Notion’s own published rates, verified August 2026.

Best for. Teams whose specs, notes and wikis live in Notion and who want them readable and writable from an assistant.

4. Slack, best for channel and thread context

Overview. The widely used Slack server exposes messages, channels and search. We run it and count 13 tools, which makes it one of the lightest connections on this list.

Key features. Reading channels and threads, searching public and private history, reading profiles, sending and scheduling messages, updating canvases.

Pros & cons. ✅ Small tool surface, so it costs little context to keep connected ✅ Search across history turns Slack into usable memory for an assistant ⚠️ Community-maintained, so review its permissions before connecting a real workspace ⚠️ Message sending is a write action worth gating

Pricing. Free and open source; a Slack workspace is required, verified August 2026.

Best for. Teams who want an assistant to summarise channels, find past decisions and draft updates without leaving the conversation.

5. Supabase, best for a Postgres backend

Overview. Supabase’s official server connects an assistant to your project’s database and infrastructure. We run it and count 29 tools.

Key features. Running SQL, inspecting tables and extensions, applying migrations, deploying edge functions, querying logs, fetching security and performance advisors, managing branches.

Pros & cons. ✅ Log queries and advisors make it useful for debugging, not only for reading data ✅ A read-only SQL variant exists, which is the right default ⚠️ Migration and branch tools write to real projects, so scope the token deliberately ⚠️ 29 tools is heavy if all you want is to read a table

Pricing. Free tier available; paid Supabase plans start at the vendor’s published rates, verified August 2026.

Best for. Product and engineering teams building on Supabase who want to explore data and diagnose issues from their assistant.

6. Airtable, best for a lightweight ops database

Overview. The Airtable server gives read and write access to bases, with schema inspection so an assistant understands your tables before touching them. We run it, and at 43 tools it is the heaviest connection we measured outside Peec AI.

Key features. Listing bases and tables, reading schemas, searching and filtering records, creating and updating records and fields, working with interface pages.

Pros & cons. ✅ Schema inspection first means fewer wrong writes ✅ Interface-level reads work on bases where you only have interface access ⚠️ 43 tool definitions is a large context cost for one connector ⚠️ Community-maintained, so check the token scope you issue

Pricing. Free with an Airtable account; paid Airtable plans at the vendor’s published rates, verified August 2026.

Best for. Ops teams using Airtable as a light database who want records queried and updated conversationally.

7. Webflow, best for site and CMS content

Overview. Webflow’s official server lets an assistant manage site content and structure. We run it and count 31 tools, spread across CMS data and design-surface operations.

Key features. Reading and updating CMS collections and items, managing pages and settings, working with components and styles, handling assets and fonts.

Pros & cons. ✅ Covers content and structure, so an assistant can draft an item and place it ✅ OAuth per site, which keeps permissions narrow ⚠️ The design-surface tools need the Webflow Designer open, which surprises people ⚠️ Publishing stays a deliberate human step, by design

Pricing. Free with a Webflow account; site plans at Webflow’s published rates, verified August 2026.

Best for. Marketing and web teams who draft or update site content programmatically.

8. Customer.io, best for lifecycle segmentation

Overview. Customer.io’s official server gives an assistant access to a marketing workspace. We run it, and at 8 tools it is the leanest server we measured, which is a design choice rather than a gap.

Key features. Reading and writing through the platform APIs, inspecting the workspace schema, listing packaged skills, checking auth status, running deletes behind an explicit call.

Pros & cons. ✅ 8 tools means almost no context cost, and the schema tool lets the assistant discover the rest ✅ Segment and campaign context grounded in live customer data ⚠️ The thin surface pushes more work into prompts than into tool calls ⚠️ Write and delete paths deserve a dry run first

Pricing. Free with a Customer.io account; platform plans at the vendor’s published rates, verified August 2026.

Best for. Lifecycle and marketing teams who want segmentation and campaign context pulled from live behaviour.

9. Peec AI, best for AI search visibility

Overview. Peec AI tracks how a brand appears across AI engines, and its server exposes that data to an assistant. We run it, and at 73 tools it is the largest connection we measured anywhere.

Key features. Brand mentions, citations and share of voice over time, cited domains and pages, per-prompt and per-engine breakdowns, brand perception attributes.

Pros & cons. ✅ The only server here that answers “how do AI engines describe us” ✅ Granular enough to attribute a citation to a specific page and engine ⚠️ 73 tool definitions is the heaviest context cost here, so connect it only for that work ⚠️ Its metrics need care: retrieval and citation rates are averages, not percentages

Pricing. Paid product; Peec AI plans at the vendor’s published rates, verified August 2026.

Best for. Marketing teams measuring how they show up in AI answers, alongside classic search reporting.

10. Zapier, best for breadth across apps with no native server

Overview. Zapier’s server is a middleware layer rather than a single integration: one connection fronts actions across a very large app catalogue. We did not run it, so this entry stays factual.

Key features. Exposing Zapier actions across thousands of connected applications, with OAuth against your account and per-action configuration.

Pros & cons. ✅ Reaches tools with no native MCP server, which is most tools ✅ One connection instead of many keeps the tool surface predictable ⚠️ An extra layer sits between assistant and app, so errors are harder to read ⚠️ Not measured in our stack, so no field verdict here

Pricing. Free tier available; paid Zapier plans at the vendor’s published rates, verified August 2026.

Best for. Teams who need reach across many applications more than depth in any one of them.

Sales, marketing and GTM

• HubSpot (official, OAuth). Reads and acts on CRM records: contacts, companies and deals. Best for revenue teams whose pipeline already lives in HubSpot and who want to query it in natural language.
• Salesforce (official, OAuth). Covers the DX server for development tasks and hosted servers that connect an assistant to org data and actions. Best for enterprise teams needing an assistant inside a governed Salesforce org.
• Attio (official, OAuth). Full CRM coverage across deals, tasks, lists, people, companies and notes. Best for teams who run on Attio and want CRM operations without raw API calls.
• Apollo (official, API key). Searches people and companies matching an ICP, enriches contacts, creates records, enrolls prospects. Best for the sourcing step upstream of outreach execution.
• Intercom (official, OAuth). Reads conversations, contacts and support data. Best for support and success teams who want ticket triage and customer context summarised on request.

Developer and infrastructure

• Filesystem (reference implementation, local scope). Scoped read and write access to local files, the backbone of most coding setups. Best for anyone pointing it at one project directory rather than a whole machine.
• Playwright (Microsoft, local scope). Browser control for navigating pages, filling forms and reading content. Best for tasks that need a real browser, including testing and pulling data from sites with no API.

Productivity, data and payments

• Linear (official, OAuth). Creates issues, updates status and queries projects. Best for engineering and product teams who file and triage work without leaving the assistant.
• n8n (official, token). Connects to and runs n8n workflows programmatically. Best for teams who already automate in n8n and want those workflows callable as tools.
• Make (official, token). Triggers on-demand scenarios and returns their results as structured data. Best for teams with an existing Make library they want an assistant to invoke.
• Stripe (official, API key). Exposes payments, customers and billing data. Best for support and finance workflows where you would otherwise be clicking through the dashboard.

…for outbound sales and GTM teams

Connect La Growth Machine for building audiences, launching sequences and handling replies, and Apollo upstream for sourcing and enrichment. If you want the assistant to do more than call tools, and to actually run a sequence of steps on its own, the pattern to look at is AI agents for sales automation rather than another connector.

…for revenue and lifecycle marketing

Connect Customer.io for segmentation grounded in live behaviour and Peec AI for how AI engines describe you. This pair is unusually well balanced on context cost: 8 tools against 73, so keep Peec AI connected for reporting sessions and disconnect it when you are doing something else.

…for engineering teams shipping with AI

Connect Filesystem so the assistant can read the project, GitHub for pull requests and issues, and Supabase if your backend lives there. That is roughly the ceiling before tool definitions start crowding out working context, which is why deferred tool loading now ships by default in Claude Code and Cursor.

…for solo founders and small teams

Start with one server tied to the tool you open most, usually Notion or Slack, and add a second only when you hit a job the first cannot do. Both are light: 28 and 13 tools respectively. A solo stack of five servers spends more context on tool descriptions than on your actual question.

Check who maintains it

Vendor-published servers track the product’s own API and get patched when it changes. Third-party servers vary, and a fork of a fork is a supply chain you did not choose. Two servers we run, Slack and Airtable, are community-maintained, and we scope their tokens more narrowly.

Check how it authenticates

The 38% to 41% figure comes from two independent 2026 measurements, an audit of 518 servers in the official registry and a broader scan of roughly 12,000 public servers, verified August 2026. Prefer OAuth or a scoped token. Grant the narrowest scope that works, and make sure write actions require confirmation rather than firing silently.

Check what it costs you in context

Ask how many tools the server loads before connecting it. The range we measured runs from 8 to 73, the difference between a server you leave connected and one you turn on for a session. Stack two or three of the heavy ones and tool descriptions start competing with your actual question for room.

Method 1: Claude Code (CLI or config)

From the terminal, add a server with claude mcp add. For a local server, pass its command; for a remote server, use --transport http and the URL. To connect the La Growth Machine server:

claude mcp add --transport http lgm https://mcp.lagrowthmachine.com/

Claude Desktop and Cursor use the same idea through a JSON config entry, with a command or URL per server.

Method 2: Customize (Claude.ai and Claude Cowork)

In the browser or in Claude Cowork, open the Customize panel. Nothing to edit by hand, and two options:

• Add a server from the marketplace of connectors: pick a listed server, authenticate, approve it.
• Add a custom server: paste the URL and its authentication for anything not in the directory, including your own.

To install the La Growth Machine server together with its GTM skills in one step:

curl -fsSL https://raw.githubusercontent.com/LaGrowthMachine/gtm-system/main/install.sh | sh

That pulls the open-source gtm-system repository, which holds both. For building your own, see our guide on how to build and connect an MCP server; for the difference between MCP and a plain API, see MCP vs API.

What is the best MCP server?

It depends on the job. For outbound sales and GTM, La Growth Machine; for code, GitHub; for team knowledge, Notion; for anything without a native server, Zapier. Start with the server attached to the tool you open most, and add a second only when you hit its limits.

Are MCP servers safe to install?

Vendor-published servers are generally safe when you scope the token. Community servers vary, and public scans in 2026 put between 38% and 41% of public servers at no authentication at all, verified August 2026. Review permissions and prefer servers that ask for confirmation before writing.

Which MCP servers are official?

Most on this list: HubSpot, Salesforce, Attio, La Growth Machine, Apollo, Intercom, Customer.io, Peec AI, GitHub, Supabase, Notion, Linear, Webflow, Zapier, n8n, Make and Stripe. Slack and Airtable are community-maintained, and Filesystem is a reference implementation.

Do I need to code to use an MCP server?

No. Connecting an existing server is one config entry in Claude Code or one connector in the Customize panel. Code is only needed if you build your own. If you want packaged workflows on top of the connection, the best Claude Code skills cover that layer.

How many MCP servers can I connect at once?

Technically many, since a host runs one client per server. Practically, two or three. Past five, tool definitions crowd the context window, which is why Claude Code and Cursor both moved to loading tool schemas on demand.