TL;DR: As of July 2026, Salesforce, HubSpot, Gong, Zoom, Atlassian (Jira), Linear, Intercom, Slack, Notion, Fireflies, Fathom, Granola, and Pendo all ship official MCP servers. Zendesk has announced one but hasn’t shipped it. Freshdesk’s is in gated beta. Every claim below is verified against a primary source with a link — because most of what’s written about MCP availability is guesswork.
I’ve spent two decades building integration platforms — Boomi, Salesforce Flow, the Slack Platform — and I’ve never seen a protocol go from spec to shipped-by-every-major-vendor as fast as MCP. I’ve also never seen so much confident misinformation about what actually exists.
Most “best MCP servers” roundups hedge every entry with “status unconfirmed” or, worse, describe servers that were never announced. So we did the boring thing: we checked every vendor’s docs, changelogs, and press releases. Here’s the state of the revenue stack.
The tracker: official MCP servers, verified July 2026
| Tool | Official MCP server | Access | Read/write | Source |
|---|---|---|---|---|
| Salesforce | Shipped — GA April 29, 2026 (Enterprise Edition+) | Remote, Salesforce-hosted; enabled in Setup → API Catalog | Read + write, runs as the authenticated user with CRUD/FLS/sharing enforced; a read-only server is available | GA announcement |
| HubSpot | Shipped — GA April 2026 | Remote: mcp.hubspot.com; OAuth per user | Read + write (contacts, companies, deals, tickets); respects user permissions | Changelog |
| Gong | Shipped — announced Oct 2025, server docs published May 2026 | Remote, connected via a Gong integration | External agents query call data; Gong is also an MCP client (AI Briefer pulls external context) | Press release, docs |
| Zoom | Shipped; expanded May 2026 | Remote: mcp.zoom.us; OAuth via a Zoom Marketplace app | Meetings, recordings, transcripts, chat; some write actions | Dev docs |
| Zendesk | Announced only — server promised for “summer 2026” early access; an MCP client is in EAP today | — | — | Relate 2026 |
| Intercom | Shipped | Remote: mcp.intercom.com; OAuth or API token; US-hosted workspaces only | Read-only — six tools (conversations, contacts, search) | GitHub |
| Freshdesk | Gated beta — Enterprise-plan EAP by request | Vendor-provided; API-key auth | Read + write (create/update tickets, reply) | EAP article |
| Atlassian (Jira) | Shipped — Rovo MCP Server GA February 2026 | Remote, Atlassian-hosted; OAuth 2.1 or API tokens | Read + write (search, create and update work items and pages) | GA blog |
| Linear | Shipped — one of the earliest remote MCPs; initiatives and milestones added Feb 2026 | Remote: mcp.linear.app; OAuth 2.1 | Read + write, 25+ tools | Docs |
| Slack | Shipped — GA, but partner-gated: works through approved apps (Claude, ChatGPT, Perplexity, Cursor…), not arbitrary clients | Remote, Slack-hosted | Search, channel history, send messages, canvases | Help center |
| Notion | Shipped | Remote: mcp.notion.com (OAuth); open-source local server too | Read + write (pages, databases, comments) | Docs |
| Fireflies | Shipped | Remote: api.fireflies.ai/mcp; API key or OAuth connector | Read (transcripts, summaries, action items) | Docs |
| Fathom | Shipped — official docs at developers.fathom.ai | Remote: api.fathom.ai/mcp | Read (meetings, transcripts, summaries) | Docs |
| Granola | Shipped — launched February 4, 2026; enterprise access is admin-gated | Remote, via Claude/ChatGPT connectors; OAuth | Read (meetings, notes; full transcripts on paid plans) | Launch post |
| Pendo | Shipped — GA, all paid plans, admin-enabled | Remote, Pendo-hosted; OAuth | Read-only analytics (usage, segments, visitor metadata) | Product page |
| Asana | Shipped — V2 server; note the V1 beta endpoint was shut down May 11, 2026 | Remote: mcp.asana.com/v2/mcp; OAuth | Read + write (tasks, projects) | Docs |
| ClickUp | Shipped — public beta, all plans | Remote: mcp.clickup.com/mcp; OAuth | Read + write | Docs |
| Snowflake | Shipped — managed server in public preview (Cortex Analyst/Search) | Remote, Snowflake-managed | Read/query | Blog |
| Databricks | Shipped — managed servers in public preview as of July 2, 2026 | Remote, Databricks-hosted; Unity Catalog permissions enforced | Read/query (UC data, Genie Spaces, AI Search) | Docs |
Verification note: “Shipped” means we found live vendor documentation or a dated GA announcement. Where a vendor’s docs are login-gated (Gong’s endpoint details, for instance), we say what we could verify and no more. If you find something that’s changed, tell us — we’ll update the table and the date stamp above.
Which AI assistants can actually connect?
Claude supports remote MCP servers across paid tiers via custom connectors in claude.ai and the desktop app, plus local and remote servers in Claude Code. ChatGPT supports remote MCP through connectors and developer mode (Pro/Plus/Business/Enterprise), including write actions. Gemini is the odd one out: the Gemini CLI and Gemini Enterprise support custom MCP servers, but the consumer Gemini app didn’t support arbitrary custom connectors as of this writing. The spec itself is moving fast too — the current finalized version is 2025-11-25, and the next release (stateless core, extensions framework, hardened OAuth) is locked for July 28, 2026.
The part nobody puts in the roundup: N servers is a tax
Here’s what we’ve learned running MCP in production. Connecting servers one-by-one works, but it has two costs that grow with every tool:
1️⃣ Token overhead. Every connected server loads its tool definitions into the model’s context before you type a word — roughly 1,000 tokens per tool in practice. A documented five-server, 58-tool setup consumed ~55,000 tokens per session before the conversation started. And every raw tool result (a 50,000-token Gong transcript, say) is billed as input again on each subsequent turn.
2️⃣ Security review overhead. Each server above is a separate OAuth grant, a separate data-residency answer, a separate vendor review. Ten tools means ten of everything.
This is why we built Noded’s MCP the way we did: your whole customer stack — the CRM, the call recorders, the ticketing, the docs — flows into one Customer Context Graph, and Claude, ChatGPT, or Gemini connects to one server that already knows how the pieces relate. One OAuth grant, one security review, and the model reads an organized account story instead of eighteen raw APIs. To be clear about the tradeoff: Noded’s server is built for customer context — if you need to administer Salesforce or push code to Jira from your AI client, you’ll still want those vendors’ own servers alongside it.
FAQ
What is an MCP server?
An MCP (Model Context Protocol) server exposes a tool’s data and actions to AI assistants in a standard way. Instead of a custom integration per AI client, a vendor ships one server and any MCP-capable client — Claude, ChatGPT, Cursor, and others — can connect.
Which CRM has the best MCP support right now?
Salesforce and HubSpot both ship GA, remote, OAuth-secured servers with read and write support. Salesforce requires Enterprise Edition or above; HubSpot’s is at mcp.hubspot.com.
Does Zendesk have an MCP server?
Not yet, as of July 2026. Its MCP client is in early access, and its server was announced at Relate 2026 for “summer 2026” early access. We’ll update this page when it ships.
Do I need a separate MCP server for every tool?
Only if you connect them directly — and each adds token and review overhead. An aggregation layer like Noded exposes the whole stack through a single server. See everything Noded connects.