MCP server: use TR-AI from Claude, ChatGPT and Cursor

MCP server: use TR-AI from Claude, ChatGPT and Cursor

Connect AI assistants to TR-AI with the Model Context Protocol to search your knowledge base, draft answers, read metrics and add articles.

Growth Pro Open in TR-AI4 min read

TR-AI is an MCP server. Any assistant that supports the Model Context Protocol over HTTP can use it as a set of tools, signed in with a TR-AI API key or with OAuth (sign in to TR-AI in the browser and approve). The server URL and ready-to-paste setup for Claude Code, Claude Desktop and Cursor are in Settings → API, under “MCP server”.

ToolWhat it doesAPI key scope
search_knowledgeSearches the knowledge base TR-AI answers fromread
list_recent_resolutionsRecent tickets with outcome, confidence and why they were heldread
get_metricsThe same reports the AI analyst uses: summary, trends, breakdowns, held reasons, AI cost, quality, integrationsread
list_kb_gapsTopics your knowledge base doesn’t answer, with draft titlesread
list_known_issuesKnown product issues your team is working on: stage, workaround, ETA, customers affectedread
get_ticket_answerHow TR-AI would answer a question (a dry run — nothing is sent)resolve
create_kb_articleSuggests an article; an admin approves it on the KB health page before it is usedkb
run_regressionStarts a regression test runwrite
  1. 1
    Create a key
    In Settings → API keys, create a key with only the scopes the assistant needs — “read” is enough for questions and analysis.
  2. 2
    Add the server to your assistant
    Copy the snippet for your tool from the MCP server card and replace YOUR_TR_AI_API_KEY with the key. For Claude Code it is one command; Claude Desktop and Cursor take a small JSON block.
  3. 3
    Ask
    For example: “Using TicketResolve, which topics should we write articles for first? Draft the top one and add it to the knowledge base.”
Tips & tricks
  • Use separate keys per person or tool so you can revoke one without breaking the others; every tool call is recorded in the API audit log with the key that made it.
  • Tool calls count toward your plan’s API rate limit; get_ticket_answer uses the lower “resolve” limit because it runs the full engine.
  • Articles an assistant adds through MCP wait under “Suggested articles” on the KB health page until an admin approves them — an assistant that reads ticket text could be steered by what a customer wrote. The “kb” permission is not ticked by default on the sign-in page.
  • Clients that only support OAuth (ChatGPT connectors, Claude.ai custom connectors): add just the URL. The client registers itself, you sign in to TR-AI, pick the workspace and tick the permissions (same names as API key scopes). Only owners and admins can approve. Access tokens last 1 hour and refresh automatically; a connection unused for 30 days must be approved again.
  • Settings → Connected AI apps lists every OAuth connection with who made it; Revoke ends it at once. Connections also stop when the person who made them is removed from the workspace, drops below admin, or no longer meets the two-factor policy.