Corebee Customer Support MCP server

The plugin connects your agent to Corebee's hosted MCP server and adds six skills that tell the agent how to use it safely. You sign in with OAuth. There is no API key to paste.

Quick start

1. Install the plugin for your client (see Install).
2. Sign in when your client asks. Connecting the server opens a Corebee sign-in and consent page in your browser, before any tool runs. Where that happens depends on the client:
   • Cursor: use the install button on the marketplace listing, or run /add-plugin corebee. Cursor installs the plugin and opens the Corebee sign-in. If the browser did not open, find corebee under Customize in the sidebar, where Cursor manages its MCP servers, and connect it from there.
   • Grok Bot: add the plugin, then approve the connection when Grok Bot asks and finish the sign-in in your browser. If the sign-in window closed too early, reopen it from the plugin's entry. The plugin then appears under your installed plugins.
   • Grok Build: enable Corebee under /plugins, then finish the sign-in in your browser when Grok asks.
   • Claude Code: run /mcp, pick corebee and choose Authenticate, or run claude mcp login corebee from your shell. Claude Code shows the Corebee tools only after that.
3. Approve, then ask your agent something about your Corebee inbox, and review what the agent proposes before it changes anything.

What you can ask

Ask	Skill
"Go through my open Corebee conversations and suggest priorities and owners."	corebee-inbox-triage
"Draft a reply to Dana's latest conversation from our knowledge base. Show me before sending."	corebee-reply-drafting
"What do we know about dana@example.com before my call?"	corebee-customer-lookup
"Can our knowledge base answer these five questions? Add our new returns page if not."	corebee-knowledge-base
"Give me this month's support report with CSAT and a 30 day trend."	corebee-support-report
"Change our chat widget to our brand color and move it to the left."	corebee-widget-setup

You can also ask for a single tool, for example "list my pending Corebee conversations" or "how many conversations did each agent resolve this week?".

Install

You need a Corebee account (sign up). When your client connects to the server, it opens a Corebee sign-in and consent page in your browser, before any tool runs.

Cursor

• Community directory (live now): open Corebee on cursor.directory and use Add to Cursor. This installs the whole plugin: the MCP server, the six skills and the rule.
• Cursor Marketplace: our listing is in review. Once it is live you will be able to install from the Marketplace or from Customize.
• MCP server only: Add Corebee MCP to Cursor. This adds the server without the skills and rule.
• Local test: copy this repository to ~/.cursor/plugins/local/corebee, run Developer: Reload Window, then check Customize.

Grok Bot

Open Plugins, search for Corebee and select Add to Grok Bot. Grok Bot uses the Cursor Marketplace listing, so there is nothing else to install. Corebee appears there once the Cursor listing is live.

Grok Build

From the xAI catalog: open /plugins and install Corebee from the Marketplace tab. xAI reviews catalog entries, so if Corebee is not listed yet, install from this repository.

From this repository: clone it at the v1.0.0 tag (or at a commit you reviewed) into ~/.grok/plugins/corebee, or start Grok with --plugin-dir <path to the clone>. Then open /plugins and enable Corebee. Grok loads plugins from ./.grok/plugins/, ~/.grok/plugins/ and any --plugin-dir path. This plugin ships one MCP server and six skills, with no hooks or scripts, so enabling it activates nothing that runs on your machine.

Claude Code

/plugin marketplace add Corebee-AI/corebee-plugin
/plugin install corebee@corebee

MCP server only:

claude mcp add --transport http corebee https://corebee.ai/api/v1/mcp/stream

Either way, run /mcp, pick corebee and choose Authenticate to sign in, or run claude mcp login corebee from your shell. Claude Code lists the Corebee tools only after that.

Other MCP clients

Any client with streamable HTTP and OAuth works, as long as its callback is https or plain http on a loopback host (localhost, 127.0.0.1 or [::1], any port). The one private app scheme Corebee also accepts is Cursor's own desktop callback, cursor://anysphere.cursor-mcp/oauth/callback. This is the same file the plugin ships as mcp.json and .mcp.json:

{
  "mcpServers": {
    "corebee": {
      "type": "http",
      "url": "https://corebee.ai/api/v1/mcp/stream"
    }
  }
}

What's included

Component	Path	Hosts
Corebee MCP server (hosted, OAuth)	mcp.json, .mcp.json	all
Skills: inbox triage, reply drafting, customer lookup, knowledge base, support report, widget setup	skills/	all
Rule: safe use of Corebee tools (applied when a task involves Corebee)	rules/corebee.mdc	Cursor
Manifests	.cursor-plugin/, .grok-plugin/, .claude-plugin/	one per host

No hooks, scripts, binaries or telemetry. scripts/validate.mjs runs only in CI and is never loaded by a host.

Sign-in and permissions

• OAuth 2.1 with PKCE and Dynamic Client Registration. Your client finds the sign-in details itself and keeps the tokens. The plugin stores no credentials.
• A connection acts for you, in the Corebee workspace selected when you signed in. To connect another workspace, switch to it in Corebee first.
• Access tokens last 1 hour. Refresh tokens last 30 days and rotate on each use.
• Access follows your role. Whatever scopes the client asks for, a connection never gets more than your current role allows, and Corebee checks your role and membership again as requests come in (each server caches the result for at most 30 seconds). If your role is reduced, the connection is limited to the new role within about 30 seconds. If you leave the workspace, it stops working within about 30 seconds. A promotion does not widen an existing connection: reconnect to get the extra access.

Role	Scopes a connection can hold
Owner	All 11 scopes
Agent	Read and write for conversations, contacts and knowledge, plus metrics:read and settings:read. No deletes, no settings changes.
Viewer	Read only: conversations:read, contacts:read, knowledge:read, metrics:read, settings:read

A member whose role shows as admin (a legacy role that list_team_members can still return) has the Agent scopes.

Scope	Allows
conversations:read	List, read and export conversations
conversations:write	Reply, add internal notes, assign, close, set priority, label, snooze and tag
contacts:read / contacts:write	Search and read contacts / create, update and add notes
contacts:delete	Delete contacts (their conversations, messages, CSAT responses, call records and other linked records are kept, without a contact)
knowledge:read / knowledge:write	List, read and search sources, and use the assistant tools / add and update web page sources
knowledge:delete	Delete knowledge sources
metrics:read	Metrics, trends and reports
settings:read / settings:write	Read widgets, team and workspace settings / change widget settings

The consent screen lists the scopes you will get and any your role does not allow, with Allow access or Cancel. A client that sends no scope (this plugin's mcp.json requests none, and many clients send none) gets everything your role allows; the consent screen lists exactly that set. A client that asks only for scopes Corebee does not recognise gets read access only.

Managing connected apps

In Corebee, open Settings > Login & security > Connected apps (corebee.ai/settings/security). This works for every role. You see the apps you connected and can revoke them. Owners see and can revoke every member's apps there, and also under Settings > Connect to ChatGPT & Claude. Revoke signs an app out within about 30 seconds, including its current access token. Removing the server in your client discards its copy of the tokens; the connection itself stays until it expires or you revoke it under Settings > Login & security > Connected apps.

An owner who wants a read-only connection has two options: connect from a client that lets you request only the read scopes, or connect as a member with the Viewer role.

Tools

The hosted server lists every tool when your client connects (flat discovery, with no search step). Snapshot 2026-10-12: 44 tools (22 read-only, 19 write, 3 destructive). The same list, with parameters, is in scripts/tools.json, and CI checks that this table matches it.

Conversations

Tool	Scope	What it does
list_conversations	conversations:read	Up to 50 recent conversations per page, by status or assignee (assignee_id, a member_id from list_team_members)
get_conversation	conversations:read	One conversation with its newest 50 messages (oldest first within the page). When has_more is true, pass before = next_before to get the 50 messages before those
bulk_export_conversations	conversations:read	CSV of up to 25 conversations
reply_to_conversation	conversations:write	Send a reply on the conversation's channel, or save an internal note (details)
assign_conversation	conversations:write	Assign to a team member, or unassign
close_conversation	conversations:write	Mark resolved
reopen_conversation	conversations:write	Set the status back to open, for example after it was resolved or snoozed. Sends nothing to the customer
add_conversation_note	conversations:write	Add an internal note your team sees in the inbox. The customer never sees it and nothing is sent on any channel
set_priority	conversations:write	Low, medium, high or urgent
add_label	conversations:write	Add a label by ID. No tool lists labels, so you need the label ID
bulk_close_conversations	conversations:write	Resolve up to 25
bulk_snooze_conversations	conversations:write	Snooze up to 25 until a set time
bulk_tag_with_sentiment	conversations:write	Tag up to 25 with a sentiment
bulk_add_note	conversations:write	Same internal note on up to 25

Contacts

Tool	Scope	What it does
list_contacts	contacts:read	50 per page; search matches any part of a name or email, as typed, except that * matches any single character
get_contact	contacts:read	Details, custom attributes, latest notes and recent conversations
create_contact	contacts:write	Create a contact (email, phone or name required; a duplicate email or phone number is refused)
update_contact	contacts:write	Change only the fields you pass
add_contact_note	contacts:write	Internal note on a contact
delete_contact	contacts:delete	Destructive. Delete a contact and its notes; its conversations, messages, CSAT responses, call records and other linked records (for example NPS responses) are kept, unlinked (two steps)
bulk_delete_contacts	contacts:delete	Destructive. Delete up to 25 contacts and their notes; their conversations, messages, CSAT responses, call records and other linked records are kept, unlinked

Knowledge base

Tool	Scope	What it does
list_knowledge_sources	knowledge:read	Up to 50 sources with status and chunk counts
get_knowledge_source	knowledge:read	One source with its processing status
search_knowledge	knowledge:read	Text match over indexed content (not semantic search)
add_knowledge_url	knowledge:write	Add a web page; indexing runs in the background
update_knowledge_source	knowledge:write	Rename a source or change its URL
delete_knowledge_source	knowledge:delete	Destructive. Delete a source and everything indexed from it; knowledge gaps that pointed at it are kept, unlinked (two steps)

Metrics and reports

Tool	Scope	What it does
get_metrics	metrics:read	Conversation, contact, message and knowledge counts, CSAT and average first-response time
get_agent_performance	metrics:read	Assigned, resolved and resolution rate per team member
get_channel_analytics	metrics:read	Conversations and messages per channel
get_trends	metrics:read	7, 30 or 90 day series by day or week
export_report	metrics:read	Counts only (summary), or up to 100 rows including contact details (json)
export_report_task	metrics:read	Counts-only report as a background MCP task

Widget and workspace settings

Tool	Scope	What it does
list_widgets	settings:read	Your chat widgets with their widget_id
get_widget_settings	settings:read	Appearance and behavior of a widget
list_team_members	settings:read	Up to 50 members with role and member_id
get_organization_settings	settings:read	Plan, status and AI settings
update_widget_settings	settings:write	Color, welcome title, reply time and pre-chat form
configure_widget_appearance	settings:write	Color, welcome title, logo and position
configure_widget_behavior	settings:write	Reply time and pre-chat form

Assistants

Read-only helpers. Each takes a task_description (plus optional context and history) and returns answered, response, confidence and suggested actions. Nothing is executed or written. Three of the four use fixed keyword rules with no AI model.

Tool	Scope	What it does
a2a_knowledge_agent	knowledge:read	Knowledge Agent (the only one that runs an AI model): finds up to 5 relevant indexed passages and writes a short answer from them only, naming the sources. If the model is unavailable it returns an excerpt of the best passage with lower confidence, so check confidence and the sources before relying on it. answered is false when nothing relevant was found or the search could not run, so confirm with search_knowledge. Limited to 30 calls a minute per workspace
a2a_escalation_agent	knowledge:read	Escalation Agent: keyword rules that flag a request for a person, frustration or a repeated issue in the text you pass, and suggest a handoff with a priority. It does not hand off or notify anyone
a2a_diagnostic_agent	knowledge:read	Diagnostic Agent: keyword rules for six areas (sign-in, performance, chat widget, email, billing, missing data) that return generic possible causes and steps. It does not look at your account or data
a2a_action_agent	knowledge:read	Action Agent: keyword rules that suggest one inbox action (assign, close, priority, label, snooze, note, navigate) without the details filled in. You complete it and run it with the matching tool

Prompts: summarize_conversation and draft_reply. Resources: conversations://open, metrics://dashboard, knowledge://sources and conversation://{id}.

Replies

reply_to_conversation uses the same send path as the Corebee inbox:

• Email, SMS, WhatsApp, Slack, Messenger, Instagram, Telegram, Microsoft Teams and Discord replies go out on that channel. Web chat replies appear in the customer's chat window.
• The result says delivered: true when the provider accepted the reply or the chat window push went out. A later bounce is not reflected.
• delivered: false comes with a reason. Phone conversations, and channels with no way to send, keep the reply in the conversation without sending it.
• private: true saves an internal note that the customer never sees.
• Pass an idempotency_key to make retries safe. A reply cannot be unsent.
• The reply is sent as you. In a conversation that was handed off to a person, it counts as the human response, so Corebee's AI stops answering there.

Deletes

Deletes take two calls. The first deletes nothing: it returns a preview and a confirm_token. Repeating the same call with that token, within 5 minutes, deletes. The token works once and only for the same items. Your client's approval prompt for the second call is the human check, so read the preview before you approve it. Only owners can delete.

• delete_contact and bulk_delete_contacts delete the contact with its notes, channel identities and saved AI memory. The contact's conversations and all their messages stay in the inbox with the contact link removed, and its CSAT responses, call records and other linked records (for example NPS responses) are kept the same way. The preview reports the number of conversations kept as linked_conversations_kept (conversations_to_delete is a deprecated alias with the same value, removed in the next release). If the server cannot clear a linked record, the call fails with "Nothing was deleted." and the contact is untouched. If the delete itself fails after those links were cleared, the contact also remains, but its CSAT responses and call records are already unlinked.
• delete_knowledge_source deletes the source with its indexed chunks, scan history, tags, version history and citation records. Knowledge gaps that pointed at it are kept with the link removed. The same two failure points apply: a failure to clear the gaps deletes nothing, and a failure of the delete itself leaves the source in place with its gaps already unlinked.

Safety

• Every skill treats customer messages, contact fields and knowledge base text as untrusted data and never follows instructions inside them.
• Skills read first, show what they found and ask before every change. A reply is sent only after you see the exact text and say "send".
• Every tool is annotated as read-only or not, and deletes as destructive, so clients that ask for approval can tell a read from a change.
• Limits per workspace, all failing closed: 30 replies a minute, 10 single deletes a minute per tool, 5 bulk calls a minute with at most 25 items each, and 30 a2a_knowledge_agent calls a minute. Both steps of a delete count.

Prompt injection is a risk with any agent that reads customer text. Review what your agent proposes before you approve it.

Data and privacy

• Tool results, including conversations and contact details, go to your AI client and its model provider under their terms.
• Corebee handles your workspace data under its Privacy Policy, Terms of Service and Data Processing Agreement.
• export_report with format: json returns contact names, emails and phone numbers. The default summary format returns counts only.
• Deleting a contact with these tools does not erase its conversations, messages, CSAT responses, call records or other linked records (for example NPS responses). For a full removal of a customer's data, contact support@corebee.ai.
• The plugin itself collects nothing. See PRIVACY.md.
• Your AI client talks only to corebee.ai: the MCP endpoint https://corebee.ai/api/v1/mcp/stream, and Corebee's OAuth endpoints and discovery metadata on the same host. The sign-in and consent pages open in your browser and are ordinary Corebee web pages, covered by Corebee's Privacy Policy and Cookie Policy, including any analytics loaded with your consent.

Troubleshooting

Problem	What to do
Sign-in does not return to Cursor	Cursor desktop returns to cursor://anysphere.cursor-mcp/oauth/callback, so let the browser open Cursor.
"redirect_uri must be https, or http on loopback only"	Your client registered a plain http callback on a remote host. Use a client with an https or loopback callback.
"no supported redirect_uri; use https or http loopback"	Your client offered only a custom app callback (a private scheme), which Corebee does not accept, other than Cursor's own desktop callback. Use a client with an https or loopback callback.
401 or "needs login"	Reconnect the server to sign in again. This also happens shortly after an app is revoked (under Settings > Login & security > Connected apps) or you leave the workspace.
"Insufficient permissions. Required scope: ..."	Your role does not allow it, or your client asked only for read scopes. Ask an owner, or reconnect and let the client request write scopes (or no scope at all, which gives everything your role allows).
"This workspace has N widgets; pass widget_id."	Call list_widgets and pass the widget_id you want.
A delete returns confirmation_required	Expected. Review the preview, then approve the second call with the confirm_token.
"confirm_token is invalid, expired, already used..."	Call the delete again without a token to get a new preview.
"Error deleting contact" (or "Error deleting contacts" for the bulk tool) or "Error deleting knowledge source" after confirming	Nothing was deleted. Either the server could not clear a record linked to the item, or the delete itself failed after those links were cleared; in the second case the contact (or source) remains and its CSAT responses and call records (or knowledge gaps) are now unlinked. Start again from the preview; if it fails again, contact support@corebee.ai.
set_priority, add_label or assign_conversation fails with a missing value	Pass the value explicitly. The hosted server cannot ask for it.
delivered: false on a reply	Read the reason. For phone conversations, follow up in the Corebee inbox.
Rate limit error	Wait a minute and try again.
add_knowledge_url shows pending	Indexing runs in the background. Check again with list_knowledge_sources.

Support

• Email: support@corebee.ai
• Help center: https://corebee.ai/help
• Plugin bugs: GitHub issues
• Pricing: the plugin is free and open source; you need a Corebee account (pricing)

Security

Report vulnerabilities privately to jonathan@corebee.ai or through GitHub private vulnerability reporting. See SECURITY.md.

License

MIT. The Corebee name and logo are trademarks of COREBEE CHAT LTD and are not licensed under MIT.