Gong MCP Server: Setup, Tools & Real Limits
The official Gong MCP server brings account, deal, and brief insights into AI clients, but its read-only scope, permissions, and credit model matter.
A Gong MCP connection lets an AI assistant ask focused questions about Gong accounts and deals or generate structured briefs from calls and emails. The official Gong MCP server is the production default: it is read-only, uses OAuth, respects user access, and returns synthesized insights—not raw transcripts. Its main catches are paid-seat requirements, Gong credit consumption, and limited tool coverage.
Table of contents
- What Gong MCP actually is
- What exists today
- How to set up Gong MCP
- Tools you get
- Permissions and blast radius
- Real limits
- When not to use it
What Gong MCP actually is
Gong MCP is a controlled bridge from an MCP-compatible client—such as an AI assistant or embedded agent—to Gong’s revenue intelligence. Once connected, the client can ask about one account, ask about one CRM deal, or request a multi-section brief for an account, deal, or contact.
The important distinction is what comes back. Gong’s server analyzes linked calls and emails, then returns an AI-generated answer or brief; it does not expose raw call transcripts, email bodies, or activity lists. Gong’s server overview also says private calls are excluded and the service cannot create, update, or delete Gong or CRM records.
That makes this an insight interface, not a general Gong database connector. A useful request is “What objections surfaced on this deal in the last 30 days?” A poor one is “Export every transcript this quarter.” MCP Directory’s capability index separates question-answering tools from raw-data and action tools.
What exists today
There is now a first-party Gong MCP server, and it should be the default for production. Gong hosts it at a documented remote endpoint, owns its OAuth flow, and exposes three supported, read-only tools. This is materially safer than sending Gong API keys to an arbitrary local wrapper.
The official service is recent, so expect a narrow surface rather than full API parity. Gong’s MCP introduction promises questions, summaries, and AI-generated insights—not every Gong object or admin function. See MCP Directory’s official collection when vendor ownership is required.
The open-source kenazk/gong-mcp repository exposes list_calls and retrieve_transcripts using API access-key credentials. It can serve an experiment needing raw call metadata or transcripts, but shifts patching, credential storage, dependency review, logging, and uptime to you. I would not treat a small community repository as a production substitute for Gong’s managed service.
| Option | Best fit | Main trade-off |
|---|---|---|
| Official Gong MCP server | Supported account/deal Q&A and briefs | Read-only, narrow, and credit-consuming |
| Community API wrapper | Lab work requiring calls or transcripts | You own security, hosting, and maintenance |
| Direct Gong API script | Deterministic exports and batch processing | More engineering, less conversational flexibility |
How to set up Gong MCP
The supported setup is an OAuth integration created by a Gong technical administrator, followed by client-side registration of Gong’s remote endpoint. You need a paid Gong seat and sufficient Gong credits; collaborator seats cannot connect.
- In Gong, open Admin center → Settings → Ecosystem → API.
- Under Integrations, choose Create integration → MCP integration.
- Name it and select Personal access for an interactive assistant or Shared access for an organization-wide backend agent. Choose carefully: Gong says this setting cannot be changed after submission.
- Add the exact redirect URI supplied by your MCP client. Each URI goes on its own line.
- Save, then copy the generated client ID and client secret into the client’s OAuth configuration.
- Set the remote server URL to
https://mcp.gong.io/mcp, complete sign-in, and verify that the expected three tools appear.
These steps come from Gong’s official connection guide. The server uses OAuth 2.0 authorization code flow with PKCE; dynamic client registration is unsupported. A custom agent must implement token acquisition and refresh.
There is no universal mcpServers JSON block to paste here. OAuth setup and redirect URIs vary by client, so copying a stdio configuration would be wrong. Use the client’s documentation and MCP Directory’s config generator only after confirming its remote-HTTP and OAuth fields match.
Tools you get
The official Gong MCP server exposes exactly three documented tools today. They return synthesized Gong insights over a defined time range, not raw evidence dumps.
| Tool name | What it does | Use it for |
|---|---|---|
ask_account | Answers one natural-language question about one account | Objections, risks, stakeholder concerns, or competitor mentions |
ask_deal | Answers one natural-language question about one CRM deal | Blockers, progress, commitments, or next steps |
generate_brief | Produces a structured multi-category summary for an account, deal, or contact | Reviews, executive briefings, and handoffs |
Gong’s tool documentation says the two ask_* tools work best with a concise, insight-oriented question. generate_brief uses a predefined brief and does not accept an open question. In practice, ask one decision-relevant question at a time; broad prompts cost more context and can blur the evidence.
Permissions and blast radius
Start with personal access, because it limits results to the authorizing user’s Gong permissions. Use shared access only for a reviewed reporting or backend agent that genuinely needs organization-wide visibility; Gong explicitly says shared access can reach all company Gong data.
The official server is read-only, so a careless tool call cannot edit a deal, delete a recording, or write to the CRM. The real blast radius is disclosure: an over-broad shared token can let an assistant summarize sensitive calls and emails across the organization. Personal access still inherits whatever the user can see, so review seat assignments and permission profiles before rollout. Gong’s permission-profile guide explains that profiles define what a team member can access and can vary by workspace.
For a backend connection, use a dedicated identity with the smallest practical scope, store the secret in a secrets manager, and keep production credentials out of desktop config files. Require confirmation before copying an answer into customer-facing systems. Log the user, entity, time range, tool, and destination—without unnecessarily logging the returned narrative.
Real limits
Gong MCP has four meaningful limits: it returns summaries rather than raw records, consumes Gong credits, inherits access boundaries, and covers only three insight workflows. MCP standardization does not turn it into a bulk export or write API.
- No raw evidence: transcripts, email bodies, and activity lists are not returned; private calls are excluded. This reduces context overload but prevents transcript-level quoting and independent verification.
- Credit consumption:
ask_accountandask_dealuse Gong’s ask-entity API, whilegenerate_briefuses get-brief. Gong says equivalent MCP and API requests consume the same credits, with usage driven by analyzed data volume. - Repeated work costs again: asking a similar question twice repeats analysis and consumes credits independently. Longer date ranges, broader prompts, and briefs with more open-ended sections increase consumption, according to Gong’s credit guide.
- Context still matters: a synthesized brief can be long enough to crowd an ongoing chat. Limit the date range and entity scope, request one brief at a time, and persist approved conclusions outside the model context when appropriate.
- No writes or bulk jobs: the official tools cannot update Gong or the CRM, enumerate all calls, or retrieve all transcripts. Use the public API for those supported operations instead.
If you choose a community wrapper, you inherit the Public API’s separate operational limits. Gong documents a default company quota of three API calls per second and 10,000 per day, with HTTP 429 and Retry-After; list endpoints may require cursor pagination. Its API introduction also warns that new response fields may appear without notice. Bound concurrency, honor backoff, ignore unknown fields, and never pour hundreds of transcripts into one model prompt.
When not to use it
Do not use Gong MCP for deterministic data movement, transcript export, high-volume analysis, or any workflow that must update a system. Use it when a person or agent needs a bounded, conversational insight about one revenue entity.
For a nightly warehouse sync, compliance archive, or dashboard feed, use the Gong API and an ordinary job with checkpoints, pagination, retries, and tests. For repetitive questions already covered inside Gong, use Gong’s own AI features: they avoid another client, OAuth integration, and context boundary. The MCP server is most valuable when Gong insight must appear inside an external assistant alongside other approved sources.
My rule is simple: choose Gong MCP for interactive synthesis; choose an API script for repeatable computation; choose native Gong AI when the user already works in Gong. I would avoid deploying a community transcript server merely because it exposes more data. That trade-off buys coverage by taking on credential risk, maintenance, and a much larger privacy surface. Compare alternatives in the best MCP servers guide, but favor the smallest interface that completes the job.