Gainsight (Staircase AI)

Name: Gainsight (Staircase AI) MCP Server
Availability: InStock
Author: Gainsight (Staircase AI)

Productivity2 tools

Ask a plain-English question about a customer account and get Staircase AI's answer, drawn from the emails, meeting transcripts, chats and support tickets it has ingested. Two tools in Anthropic's directory, one of which appears in no Gainsight document at all.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on 19 August 2026.

Anthropic states this reflects the level of review a connector received, not a security audit.

Connect Gainsight (Staircase AI) via MCP

https://mcp.staircase.ai/mcp

Works in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.

Gainsight (Staircase AI) Tools & Capabilities (2)

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

One of two tools is undocumented by the vendor. staircase_fetch_evidence is named in Anthropic's directory snapshot and appears in no Gainsight support article we fetched. Its inputs, outputs and the volume of source text it returns are unknown.
Two Gainsight statements are hard to reconcile. The security article says raw customer communications do not leave Staircase except as summarised intelligence; the same table lists evidence as part of the response, and a tool is named for fetching it. We could not resolve this without authenticating, and we do not authenticate.
We could not verify the tool list or any tool behaviour. The endpoint returns HTTP 401 to an anonymous handshake. No schemas, no descriptions, no safety annotations were readable. Both tools are unannotated as far as we can tell.
Gainsight publishes no tool table. The count of two comes from the directory alone, so we can neither confirm it is complete nor show that it is short.
There are no OAuth scopes. Both discovery documents return an empty scopes_supported array. Nothing in the grant distinguishes asking a question from retrieving evidence, and the consent screen can offer no choice narrower than the whole connector.
The read surface is third-party correspondence. Emails, meeting transcripts, chats and support tickets involve people outside your company who never agreed to an AI connector reading what they wrote. Gainsight notes PII-anonymised data stays protected where that setting is enabled, which implies the protection is a configuration rather than a default.
The dispatcher's reach is not enumerable. staircase_query takes free text. Nothing in the published call bounds which accounts or communications the answer may draw on, so the connector's reach cannot be sized by reading tool schemas.
Identity-provider documentation disagrees with itself. The setup article offers Okta and Gainsight SSO; the security article's table marks Okta and SAML SSO unsupported. Both fetched 2026-08-19.
A non-interactive credential exists. Personal OAuth client credentials mint a twelve-hour token with no browser login, and the client_credentials grant is not advertised in the server's own RFC 8414 metadata. The secret is displayed once and is unrecoverable.
The vendor recommends always-allow. Gainsight's setup instructions tell users to remove per-call approval, including for the tool it has not described.
Access scoping is Gainsight's claim, not our observation. No field-level permission map is published, and no article states what a query returns for data the user cannot see.
Prompts are unconfirmed. Anthropic's snapshot lists no prompt names, but the gated endpoint blocked our own prompts/list, so we cannot say whether the server serves any.
CRM data is out of scope for now. Gainsight states the underlying feature reads only the communication layer, with renewal dates and revenue not yet included.

Frequently asked questions

What does the Gainsight (Staircase AI) MCP connector actually read?

Customer communications. Gainsight describes Staircase AI as answering natural-language questions across emails, meeting transcripts, chats and support tickets, and states the underlying feature reads only that communication layer — CRM tabular fields such as renewal dates and revenue are excluded for now. So the read surface is correspondence with your customers, not typed CRM records.

What does staircase_fetch_evidence do?

Gainsight does not say. The tool is named in Anthropic's directory snapshot of 2026-08-16 and appears in no Gainsight support article we could find, including the connector's own documentation page. Its name implies it returns supporting material behind an answer, but no vendor document describes its inputs, its outputs, or how much source text it hands back.

Does the Staircase AI MCP server send raw customer emails to the model?

Gainsight's security article says raw customer communications stay inside Staircase except as summarised intelligence. The same table lists evidence among what the server returns, and a directory-listed tool is named for fetching evidence. We could not reconcile those without authenticating, which we do not do, so treat the boundary as undocumented rather than settled.

Whose permissions limit what the connector can see?

The authenticated user's. Gainsight states MCP queries respect existing Staircase permissions, that each user authenticates individually with no shared organisation-level tokens, and that cross-organisation access is impossible. It does not publish a field-level map from tools onto those permission checks, so what a query returns for out-of-scope data is undocumented.

Can the Staircase AI MCP server write to Gainsight?

No tool documented or listed writes anything. Anthropic's directory labels the connector Read and names two tools, both phrased as retrieval. Gainsight's platform creates CTAs, timeline entries and health scores in its own product, but no support article describes an MCP tool reaching any of them. We could not confirm this on the wire.

Do Staircase AI MCP credentials require a browser login?

Not always. Gainsight documents interactive sign-in through Google, Microsoft, Okta or Gainsight SSO for LLM clients, and separately documents personal OAuth client credentials that authenticate with no browser flow at all. Those credentials mint a twelve-hour token for automated clients, so a connected process can read account intelligence without a person present.

How many tools does the Gainsight (Staircase AI) MCP server have?

Anthropic's directory snapshot of 2026-08-16 names two: staircase_query and staircase_fetch_evidence. Gainsight publishes no tool list of its own, so we cannot confirm that count. Only one of the two, staircase_query, appears in any Gainsight document — inside a sample Python client — and the server is gated, so we could not enumerate it ourselves.

Sources

Connect Staircase AI to LLMs Using MCP — https://support.gainsight.com/Staircase_AI/Staircase_AI_Features/Connect_Staircase_AI_to_LLMs_Using_MCP (fetched 2026-08-19; the documentation URL Anthropic's directory publishes for this connector; source for the org-level enablement step, the client setup flows, the region switcher and the always-allow instruction) · retrieved 2026-08-19
Staircase AI MCP Security and Compliance — https://support.gainsight.com/Staircase_AI/Staircase_AI_Features/Staircase_AI_MCP_Security_and_Compliance (fetched 2026-08-19; source for the raw-communications boundary, the intelligence-and-evidence response description, the per-user permission model, the token properties and the identity-provider table) · retrieved 2026-08-19
Create Automations Using the Staircase AI MCP Server — https://support.gainsight.com/Staircase_AI/Staircase_AI_Features/Create_Automations_Using_the_Staircase_AI_MCP_Server (fetched 2026-08-19; source for the client_credentials grant, the twelve-hour token, the client-secret warning and the only Gainsight appearance of staircase_query) · retrieved 2026-08-19
Ask Staircase — https://support.gainsight.com/Staircase_AI/Staircase_AI_Features/Ask_Staircase (fetched 2026-08-19; source for the communication-layer scope, the CRM-data exclusion and the linked-reference behaviour) · retrieved 2026-08-19
Ask Staircase Slack Integration — https://support.gainsight.com/Staircase_AI/Staircase_AI_Features/Ask_Staircase_Slack_Integration (fetched 2026-08-19; checked for tool names, none found) · retrieved 2026-08-19
Staircase AI Risk Analyst and Topics articles — https://support.gainsight.com/Staircase_AI/Staircase_AI_Features/Risk_Analyst and https://support.gainsight.com/Staircase_AI/Staircase_AI_Features/Topics (both fetched 2026-08-19; searched for a description of evidence retrieval, none found) · retrieved 2026-08-19
Gainsight Staircase AI article index — https://support.gainsight.com/@api/deki/pages/=Staircase_AI%2FStaircase_AI_Features/subpages (fetched 2026-08-19; used to enumerate the ten Staircase AI feature articles searched for staircase_fetch_evidence) · retrieved 2026-08-19
support.gainsight.com/robots.txt allows crawling of article paths with a five-second crawl delay and publishes a sitemap; it disallows only operational namespaces. www.staircase.ai serves no robots.txt (404, fetched 2026-08-19), and no marketing-site page was used as a source for this page · retrieved 2026-08-19
Live RFC 9728 protected-resource descriptor — https://mcp.staircase.ai/.well-known/oauth-protected-resource/mcp (fetched 2026-08-19; empty scopes_supported, authorization_servers naming the MCP host itself; the root and path-insert forms both return a plain-text 404) · retrieved 2026-08-19
Live RFC 8414 authorization-server metadata — https://mcp.staircase.ai/.well-known/oauth-authorization-server (fetched 2026-08-19; same empty scope array, PKCE S256, dynamic registration, authorization_code and refresh_token grants only) · retrieved 2026-08-19
Live anonymous probe — 2026-08-19: an initialize POST to the endpoint returns HTTP 401 with www-authenticate naming invalid_token and the resource-metadata URL · retrieved 2026-08-19
Anthropic connector directory — https://claude.ai/directory/7b4941fa-a8fa-4a00-a726-cf692ac269ca (snapshot 2026-08-16; source for the partner tier, the Read permissions label, the two tool names, the empty prompt list and the productivity category) · retrieved 2026-08-16