TL;DR: You can connect Gong, Jira, Linear, Intercom, and (soon) Zendesk to Claude or ChatGPT today via each vendor’s MCP server. It works — but every server you add costs roughly a thousand tokens of definitions per session, a separate OAuth grant, and a separate security review, and the model still has to stitch the customer story together in-context on every question.
MCP — the Model Context Protocol — is the USB-C moment for business software: one standard port through which an AI assistant can reach your tools. For customer teams, that promise is specific and tantalizing: ask Claude “what happened with Acme this quarter?” and have it actually know — the Gong calls, the Zendesk tickets, the Jira escalation, the renewal thread.
Here’s what that looks like in practice as of July 2026, tool by tool, and what it costs.
What can I connect today?
Gong: yes. Gong announced MCP support in October 2025 and published server documentation in May 2026. You create a Gong integration to connect, after which external agents can query call data. Gong runs the protocol in both directions — it’s also an MCP client, pulling external context into its AI Briefer.
Jira: yes. The Atlassian Rovo MCP Server went GA in February 2026 — remote, OAuth 2.1, and it can read and create/update work items. Linear: yes, and it’s one of the most mature servers out there (mcp.linear.app, 25+ tools).
Zendesk: not yet. This one surprises people. Zendesk shipped an MCP client in early access — letting you plug external servers into its action flows — but its own MCP server was only announced at Relate 2026, with early access expected summer 2026. Community-built servers exist if you can’t wait, but they’re not vendor-supported. Intercom: yes, read-only, US-hosted workspaces only (six tools). Freshdesk: gated beta for Enterprise plans.
Your recorder besides Gong: probably yes. Fireflies, Fathom, and Granola all ship official MCP servers as of early-to-mid 2026. We keep a verified tracker of the whole revenue stack, updated as vendors ship.
Which assistant should the team use?
Claude supports remote MCP servers on paid tiers via custom connectors; ChatGPT supports them through connectors and developer mode. Gemini is enterprise-and-CLI only for custom servers so far. If your team is split across assistants — and most are — note that you’ll be doing this connector setup once per tool per assistant.
The three costs of the connect-everything approach
1. Tokens. Every connected server loads its tool definitions into context at the start of each session — roughly 1,000 tokens per tool. A documented five-server, 58-tool configuration consumed about 55,000 tokens before the first question. Worse, results flow through context raw: a two-hour call transcript is ~50,000 tokens, and Anthropic’s engineering team showed it gets billed through the context window twice in a typical copy-between-tools workflow. Ask three follow-up questions and you’ve paid for that transcript five times.
2. Security reviews. Each server is its own OAuth grant, its own data-residency answer, its own DPA conversation. The questions your security team will ask of every single one: How does it authenticate (OAuth 2.1 with PKCE should be the floor)? Are scopes read-only or write? Where is data hosted, and is any of it retained? Does the vendor train on it? Does access respect existing user permissions? Multiply by every tool in the stack.
3. The joins. This is the subtle one. Even with everything connected, the model gets raw, disconnected records — a transcript here, a ticket there — and has to reconstruct the account story in-context, from scratch, on every question. It usually can. But you’re paying frontier-model prices for ETL work, and accuracy degrades as context fills with unjoined raw data (the research on this is consistent — see our token-spend guide).
The aggregation alternative
The pattern that’s emerging — in Anthropic’s own engineering guidance, in Cloudflare’s “Code Mode,” and in how we built Noded — is to do the aggregation before the model, not inside its context window.
Noded connects to the stack once — Gong, Zoom, Fathom, Fireflies, Granola, Salesforce, HubSpot, Zendesk, Intercom, Jira, Linear, Slack, email — and weaves it into the Customer Context Graph: one living story per account, deduplicated and organized. Your assistant connects to one MCP server that serves curated account context instead of eighteen raw APIs. One OAuth grant. One security review (SOC 2, never trains on your data). And dramatically fewer tokens, because the summarize-and-join work happened once, ahead of time — that’s where our up-to-90%-lower token cost figure comes from (methodology here).
Where Noded’s server is not the right tool: deep tool-specific write operations. If you want your assistant to restructure Jira sprints or administer your Salesforce org, use those vendors’ servers for that — they run happily alongside ours.
FAQ
Can Claude read my Gong calls?
Yes — via Gong’s official MCP server (docs published May 2026), or via Noded, which brings Gong calls in already connected to the account’s CRM, support, and email context.
How many tokens does connecting multiple MCP servers cost?
Roughly 1,000 tokens per tool per session in definitions alone; a five-server setup was measured at ~55,000 tokens before the conversation started. Raw results then bill as input on every turn.
What will security ask before approving an MCP server?
Auth method (OAuth 2.1 + PKCE), read vs write scopes, data residency and retention, model-training policy, and whether access maps to existing user permissions. Prepare one answer sheet per server — or connect one server and answer once.