Skip to main content
SuperBot
Start free

Developer docs · MCP connector

Use SuperBot inside Claude, Codex and ChatGPT

SuperBot ships an MCP server, so you can point Claude, ChatGPT or Codex at your own workspace and ask about your support the way you would ask a colleague — what customers are complaining about, which questions the AI keeps failing, what to teach it next.

Add to Claude Add to ChatGPT
Approved in the Claude directory and ChatGPT apps · nothing to install
On this page
  1. 01Connect it
  2. 02Start with a real workflow
  3. 03What it can do
  4. 04Permissions and data
  5. 05Troubleshooting and safe use
  6. 06Questions

Connect it

Add this URL as a connector, then sign in to SuperBot when it asks and approve the permissions. Nothing to install.

Add to Claude Approved in the Claude directoryAdd to ChatGPT Approved in the ChatGPT apps
https://superbotapp.ai/api/mcp
  • ClaudeUse the directory link above, or open Settings, then Connectors, then Browse connectors and search for SuperBot.
  • ChatGPTOpen the SuperBot app in ChatGPT and select Connect, or search for SuperBot in ChatGPT's apps directory.
  • Codex CLIAdd the remote server, then log in with read-only access:
    codex mcp add superbot --url https://superbotapp.ai/api/mcp
codex mcp login superbot --scopes superbot:read

    To enable approved training writes, reconnect with --scopes superbot:read,superbot:write. The server also checks your current SuperBot workspace role.

  • Claude Code
    claude mcp add --transport http superbot https://superbotapp.ai/api/mcp

Start with a real support workflow

  • “Summarize support over the last 7 days, including CSAT sample size and the busiest channels.”
  • “Review recent escalations. Read the transcripts and cite the conversations before suggesting fixes.”
  • “Check whether we already answer shipping questions. Draft a missing answer for my approval; do not save it yet.”

What it can do

List tools return a next_cursor when more records match; your AI client passes it back to get the next page.

  • searchRead

    Full-text search across your conversations and their messages.

  • fetchRead

    Up to 500 messages and 60,000 transcript characters, with a dashboard link and truncation flag.

  • list_recent_conversationsRead

    Newest conversations, optionally filtered by status.

  • list_unanswered_questionsRead

    Conversations that may still need an answer, for triage: human escalations first, then open or pending ones. Each is a lead to check, not proof the AI failed.

  • search_knowledgeRead

    What your AI has been taught: trained answers and knowledge sources.

  • get_support_overviewRead

    Volume, open vs closed, CSAT, sentiment and channel mix over a window.

  • get_insightsRead

    Top topics and repeated questions with trends, plus volume, escalation, CSAT and AI fallback-rate changes against the previous period.

  • list_training_gapsRead

    Questions the AI could not answer well, with evidence: AI fallback replies, human escalations and 1-2 star CSAT.

  • list_knowledge_sourcesRead

    Every knowledge source your AI learns from, with type, status, entry count and last sync.

  • list_leadsRead

    Leads captured by your bot and CRM, filterable by status, source, text or date.

  • add_trained_answerWrite

    Teach your AI a new question and answer. Adds only; never edits or deletes. Side effect: your website widget can use the answer with visitors right away. Repeating the same call adds nothing. Remove an entry any time under Training in your dashboard.

Classification. Every Read tool only retrieves data and changes nothing. add_trained_answer is the only Write, and it is not a sensitive write: it moves no money, sends no messages, shares nothing outside your workspace and deletes nothing, and its single entry can be removed afterwards. No tool accepts payments.

Permissions and data

superbot:read

Covers everything above marked Read.

superbot:write

Needed only for add_trained_answer. Grant read alone and the write tool is refused. The SuperBot sign-in screen lets you untick write access to connect read-only, even when the client asks for both.

  • The connection is tied to your account and only ever reads your own workspace. No tool accepts a workspace or account id, so it cannot be pointed at anyone else's data.
  • Authorization is OAuth 2.1 with PKCE. Access tokens last an hour; refresh tokens rotate on every use. Only hashes are stored, never the tokens themselves. Grants are bound to the workspace approved at connection time; changing workspace membership requires reconnecting.
  • This connector is for your team, not your visitors — it reads the support data your widget produces. The widget itself keeps answering customers on your site as usual.
  • Disconnect it in your AI tool at any time. See our privacy policy for how support data is handled.

Troubleshooting and safe use

No results means this workspace has no matching data, not that the connector analyzed another account. Open a returned dashboard URL to verify the source. If a transcript is marked truncated, review the complete history in SuperBot. Invalid arguments return an error instead of being silently changed; an optional field left empty is treated as not set.

Errors come back as MCP tool errors with a plain-English reason, such as a missing scope, an invalid argument or an unknown id. The connector allows 120 calls a minute per account; above that it answers HTTP 429 with a Retry-After header.

If access expires, reconnect through your AI client. The server advertises a token-revocation endpoint for clients that support disconnect revocation.

SuperBot is listed in both the Claude directory and ChatGPT's apps directory. Adding the URL as a custom connector still works in clients that support it; which option you see depends on your AI client and plan.

Questions

Email [email protected] or open a conversation.

Ask your support data anything.

Connect once, then ask Claude, ChatGPT or Codex what customers asked, what the AI missed, and what to teach it next.