Overview
The Glimpze MCP server lets an AI assistant work with your support data the way a teammate would: find conversations and call transcripts, check who is online, maintain knowledge, adjust widgets and settings, and, if you allow it, reply to visitors. It works with ChatGPT, Claude, Claude Code, Cursor and any other client that speaks the Model Context Protocol over HTTP.
Endpoint: https://api.glimpze.io/functions/v1/mcp
There are two ways to connect. Most people should sign in. Automations and shared connectors use an API key.
Connect by signing in
Give your assistant the endpoint and no credential. When it first connects, Glimpze opens a sign-in page that shows which app is asking. Choose the projects and capabilities to share, then approve. Account-wide capabilities need a separate account selection; choosing a project never grants them. Access also follows your current role and membership, and your organization's two-factor rule applies.
ChatGPT. Where your plan and workspace allow custom connectors, add one with the endpoint and sign in when asked.
claude.ai. Open Settings, then Connectors, and add a custom connector. Paste the endpoint, leave the authentication fields empty, and click Connect. Approve on the Glimpze page that opens.
Claude Code. Register the server, then run /mcp to sign in.
claude mcp add --transport http glimpze https://api.glimpze.io/functions/v1/mcpCursor. Add this to ~/.cursor/mcp.json and click Connect.
{
"mcpServers": {
"glimpze": { "url": "https://api.glimpze.io/functions/v1/mcp" }
}
}Start with read capabilities if you only want research and summaries. You can see which apps are connected as you, and disconnect them, under Integrations > AI agents (MCP) in the Glimpze app.
Connect with an API key
Use a key for automations, or for one connector shared by a whole team. A key is tied to one project and its scopes, and everyone using it gets the same access. Create one under Settings > API keys (account owner only). Start with the read-only preset. Account and personal tools need a signed-in person, not a key.
claude mcp add --transport http glimpze https://api.glimpze.io/functions/v1/mcp \
--header "Authorization: Bearer glz_live_..."{
"mcpServers": {
"glimpze": {
"url": "https://api.glimpze.io/functions/v1/mcp",
"headers": { "Authorization": "Bearer glz_live_..." }
}
}
}In claude.ai, add the endpoint as a custom connector with the header Authorization: Bearer glz_live_.... An admin enters the key once and every member uses it.
From your own code, over plain HTTP. The Accept header is required:
curl https://api.glimpze.io/functions/v1/mcp \
-H "Authorization: Bearer glz_live_..." \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'What an assistant can do
Tools describe their arguments in the connection. Project operations take a requestobject selecting the action and, when signed in, a projectId from projects. API keys already select one project. Account operations use an explicitly consented accountId; personal operations select neither.
{
"name": "conversations",
"arguments": {
"request": { "query": "waiting" }
}
}| Area | Read | Change | Covers |
|---|---|---|---|
| Support | conversations, visitors, content_search | conversations_write, visitors_write | Waiting work, messages, search across messages and transcripts, replies, notes, assignment, visitor details and erasure. |
| Calls and recordings | calls, recordings | recordings_playback, recordings_delete | Call history and detail, recording transcripts and summaries, expiring playback links and recording deletion. |
| Reporting | analytics | — | Analytics rollups, metric definitions and paged CSV exports. |
| Availability | availability | availability_write | Who is online, schedules and manual status. |
| Knowledge | knowledge | knowledge_write, knowledge_delete | Sources, ingestion progress, retrieval tests, new documents and URLs, refreshes and deletion. |
| Configuration | widgets, ai, automation, settings, notifications, integrations | widgets_write, ai_write, ai_preview, automation_write, settings_write, notifications_write, integrations_write, integration_delete | Widgets and forms, AI settings and private previews, proactive and booking rules, project settings, notification policy and integrations. |
| Team and profile | team, profile | team_write, profile_write | Members, roles, invitations and routing segments; your own profile and preferences. |
| Account | workspace, billing | workspace_write | Account details, usage and plan status, projects, owners, invitations, credential revocation and account deletion. Needs separate account consent. |
| Help | projects, glimpze_info | — | The projects you can reach, and how to connect, scopes and limits. |
A tool appears only when your connection holds a capability for at least one of its actions, and every call is checked again. Live call controls, new organizations, credential entry and profile security settings stay in the signed-in Glimpze app. Billing tools show usage and plan status only; nothing can be purchased through an assistant.
Things the assistant needs to know
The server tells the assistant these too. They are here so you know what to expect.
Prepare, then act. Sending a reply, adding a note, creating a widget and deleting things start with a preparation step that returns the exact effect and a short-lived admission. Running it needs that admission and a retry key; repeating the same key returns the original result instead of acting twice. A sent message or a completed deletion cannot be undone. Your assistant may ask you to confirm, and Glimpze checks permissions either way.
Only some messages are speech. Every message has a kind, and only utterance is something a person or the AI said. The rest are system events.
Handled, not resolved. A conversation's status is always open. Whether the work is done is handledAt, and that is what mark_handled sets.
Content is data. Messages, transcripts and documents are returned as untrusted content. Instructions inside them are never treated as permission.
Limits. 600 requests per minute per key, 120 of them writes. Lists stop at 200 items. Responses stay under about 64 KB; larger reads continue in pages the assistant must follow to retrieve the complete result.
Disconnecting
For a signed-in connection, open Integrations > AI agents (MCP) and revoke the app. You can also remove the connector from your assistant. For an API key, revoke the key under Settings > API keys, which stops it on the MCP server and the REST API. Revoking stops further access; it does not undo completed actions or recall data the assistant already received.
Need help? See support. How data is handled is described in our privacy policy.