UserGist MCP
Use your AI assistant to investigate feedback, understand users, manage requests and roadmap items, write help content, and prepare targeted experiences in UserGist. The hosted Model Context Protocol (MCP) server supplies tools and evidence; your connected AI handles reasoning and writing. UserGist does not run another LLM inside the MCP server.
MCP is disabled by default and requires a configured hosted environment and an enabled workspace. The client setup guides are documentation-verified, not a claim that real hosted connections have passed acceptance testing. If AI connections shows that your workspace is not enabled, contact your administrator.
Start with a connection
- Open your workspace settings and select AI connections. Copy the MCP URL shown there.
- Add that URL to your assistant using the client setup guide. Select OAuth authentication, not an SDK write key or server API token.
- Sign in to UserGist, select one workspace and its apps, and review the capabilities. Keep Approve changes selected unless you deliberately need another mode.
- Ask your assistant to call
get_contextandlist_appsbefore working with an app.
One connection belongs to one existing workspace member and one workspace. Every app-specific action includes an explicit appId; there is no shared “currently selected app.” Public portal visitors and SDK participants cannot use these member tools.
What you can do
| Task | Example request to your assistant |
|---|---|
| Investigate feedback | “Find feedback about onboarding in this app. Cite the underlying answers and distinguish anonymous from identified users.” |
| Understand a user | “Read this subject's activity and feedback. Summarize the evidence without guessing missing facts.” |
| Manage product work | “Review these duplicates, preview a merge into this canonical request, and show follower-notification effects.” |
| Maintain a roadmap | “Move this item to shipped, changing only the linked requests I select.” |
| Publish content | “Draft a release note and help article. Show me the drafts before publishing.” |
| Target an announcement | “Create a draft in-app message using app branding. Preview its audience and schedule before activation.” |
| Diagnose delivery | “Explain why this user might not receive this experience. Do not claim or send anything.” |
| Inspect native push | “Check push readiness, then preview this campaign. Do not send a test until I name a recipient and approve it.” |
See the tool reference for the complete catalogue and supported argument shapes. Tools are filtered by the connection's grants, your current role, and rollout controls.
Choose how changes are approved
| Mode | UserGist behavior |
|---|---|
| Read-only | Reads, previews, analytics, and diagnostics are available; product writes are blocked. |
| Approve changes — default | Every product mutation, including draft edits and test sends, first produces a proposal for confirmation in the conversation. |
| Automatic | Granted actions execute without an additional UserGist approval step. Validation and current permissions still apply. |
Your AI client can require its own confirmations in every mode. When native forms are unavailable, the client is responsible for asking you to confirm; a proposal ID binds an action to its preview but does not prove human approval. Read permissions and approvals before enabling writes.
Boundaries
MCP does not expose arbitrary SQL or HTTP execution, customer impersonation, fabricated votes or responses, identity or consent changes, permanent deletion, billing purchases, team administration, credential rotation, or unrestricted transactional push. Existing provider secrets and server credentials are not returned. Creating an app can return its publishable SDK client key.
The connection does not replace an SDK integration. Anonymous and identified SDK participants, explicit aliases, and portal-origin identities remain distinct. Web push is not supported. Native push needs valid APNs/FCM configuration, a registered token, and confirmed notification receipt; credential presence alone is not delivery validation.