Navigation

Insights AI

Introduction Configure Insights AI The Insights AI assistant Connect your assistant (MCP) Export your calls
Insights AI · 10 min read

Connect your assistant (MCP)

Analyze your calls directly from your AI assistant (Claude, Claude Code, Cursor…) with the Insights AI MCP server.

View .md Open in Claude

#Why connect your own assistant?

If your team already uses an AI assistant such as Claude, you can give it access to your calls. You then ask it questions exactly like the built-in assistant, but from the tool you use every day, and you can combine your calls with your other documents and connectors.

This connection relies on MCP (Model Context Protocol), the open standard that lets AI assistants use external tools. You don't need to know anything about MCP to use it: just follow the steps below.

#What your assistant can do

Once connected, your assistant can:

  • Count and compare calls: by subject, topic, agent, queue, dialed number, outcome (answered, missed, busy), sentiment, day, week or month
  • Understand the reasons for calls and frustration from samples of call analyses
  • Assess and coach an agent against the rest of the team, including conversation KPIs (speech ratio and rate, questions, monologues, first call resolution)
  • Analyze the quality audits: score of each audit criterion and the AI justifications of the lowest scores
  • Search calls by period, direction, agent, subject, topic, dialed number or content
  • Read the full transcript and the detailed audit of a call
  • Export calls to an Excel or CSV file, with the columns you ask for, and link to the calls list of CX-Engine filtered on the calls behind an answer (see below)
  • With Expert Statistics, report figures computed from every call of the phone system: queue volumes, answer rate and waiting times, the call log, repeat callers, calls per agent and per company number, outbound activity, agent presence and who is available right now (see below)

Access is read-only: your assistant cannot change anything in CX-Engine. The only thing it can create is the call exports you ask for.

#Example prompts

  • Which subjects caused the most frustration to our customers this month?
  • How many calls did we miss last week, per queue and per day?
  • Which audit criteria fail the most, and what do the AI justifications say?
  • Prepare a coaching sheet for Julie, with her weakest audit criteria and examples of calls.
  • What do people call our sales number (+33 1 23 45 67 89) about?
  • Export September's calls on extensions 1587 and 1588 with the date, duration, agent, audit score, net sentiment score, subject and sub-subject.

#Available tools

Your assistant picks the right tools on its own: you never need to name them. They are listed here so you know what it can rely on.

Tool What it does
get_context Describes the PBX: timezone, date range of the data, agents, queues, dialed numbers, subject tree, and whether Expert Statistics is available. Called first
count_calls Counts the calls and computes metrics (durations, sentiment, frustration rate, answer rate, audit score), grouped by subject, topic, agent, queue, dialed number, outcome, sentiment or period
search_calls Lists the calls matching filters (period, direction, agent, queue, dialed number, outcome, subject, topic, sentiment, duration or free text), sorted by date, duration, sentiment or relevance
sample_call_insights Returns a sample of call summaries, reasons for call or frustration reasons, to identify recurring themes
agent_report Assesses an agent against the team: volumes, sentiment, frustration, audit score, conversation KPIs, weakest audit criteria, top subjects, best and worst calls
audit_report Reports the quality audits: score of each criterion, how often it fails, AI justifications of the lowest scores, conversation KPIs
get_call Returns one call in full: AI analysis, detailed audit and transcript
calls_list_url Returns a link to the calls list of CX-Engine, filtered on the matching calls. count_calls and search_calls also return it
export_calls Exports the matching calls to an Excel or CSV file, with the requested columns in the requested order (every column when none is given). Up to 20,000 calls
get_export Gives the status of an export and, once it is ready, its download links
find_similar_tags Finds the topics close to a phrase (e.g. "refund"), to search calls by meaning rather than exact words

Every tool accepts the same filters (period, agents, queues, dialed numbers, outcome, subjects, topics, sentiment, frustration, audit score, net sentiment score, duration, free text), so they can be combined freely.

Ask your assistant for a list or a file of calls, and it starts an export in CX-Engine with exactly the filters of its analysis. The export runs in the background: you are notified in CX-Engine (bell icon) when the file is ready, and your assistant can give you its download links if you ask. Like the other files exported from CX-Engine, they open in the app, once you are logged in, and stay available for 7 days. See Export your calls for the available columns.

Exports require the Export calls permission (crm-phone-call.export) for the token owner: without it, the export tools are not offered.

Your assistant can also add to its answers a link to the calls list of CX-Engine, filtered on the calls it relied on, so you can browse, refine and export them. Topics and meaning-based searches cannot be reproduced in the calls list: your assistant then says which filters were not applied.

#Expert Statistics

When the PBX of the token has an active Expert Statistics subscription (or trial), your assistant gets nine more tools, computed from every call of the phone system rather than only the analyzed calls:

Tool What it reports
xs_queue_report Offered, answered, abandoned (including during the welcome message), dissuaded and transferred calls, answer rate, average waiting and talking times, distribution of the answered calls by waiting time, optionally per agent
xs_trend Received and answered calls, answer rate and average waiting time per quarter-hour, hour, day, week, weekday or month, for queues or agents
xs_agent_activity Minutes in every status, availability rate and time connected to each queue, per agent (whole days only)
xs_call_log The individual calls, newest first, including the calls never analyzed by the AI: direction, caller number, dialed company number, whether an agent answered, duration, and the route through the PBX (queues, agents, IVR, voicemail, external numbers). Filters on caller or company number, direction, answered or not, agent and queue. 93 days at most, paginated
xs_caller_history The top callers of the period (calls, answered, unanswered, answer rate, talk and waiting times), plus the repeat callers and their share of the calls. For one caller number: its totals, split per queue
xs_agent_calls Per agent: inbound calls answered and unanswered, answer rate, talk time and ring time before answering, plus outbound and internal call segments
xs_did_report Per company number (DID): received, answered and abandoned calls, answer rate and talk times, optionally split per queue with the average waiting time. 93 days at most
xs_outbound Outbound activity of the whole PBX: calls made, connected calls, connection rate and talk times, per day, week, month, hour of the day or weekday
xs_live_status The current status of every agent (available, away, do not disturb, custom status, phone not connected) and how long they have been in it. Data up to 5 minutes old

The calls of xs_call_log that were analyzed by the AI carry their call id: your assistant passes it to get_call to read their transcript and analysis. The other calls have no AI analysis.

Your assistant can then combine both sources in one answer:

  • What was our answer rate on each queue last month, and what did the customers who called during the worst hours complain about?
  • At what time of day do we lose the most calls, and were enough agents available?
  • Compare the waiting times of September and October, and the frustration reasons over the same periods.
  • What happened to the call from +33 6 12 34 56 78 at 10:12? Read me its transcript if it was analyzed.
  • Who called us most this month, did they get through, and what were they calling about?
  • Which of our numbers has the worst answer rate?
  • How many outbound calls did we make last month, and at what time do we connect best?
  • Who is available right now?

The tools are offered only when the PBX has an active Expert Statistics subscription or trial, and the token owner has the Expert Statistics permission (expert-statistics.view). The get_context tool tells your assistant whether they are available.

#Step 1: create an access token

  1. In CX-Engine, select your account, then open CRM Connector > Insights AI (MCP)
  2. Click New access token
  3. Give it a name to find your way around (for example the name of your assistant), choose the PBX to analyze, then confirm
  4. Copy the displayed token: for security reasons, it will never be displayed again

The window also shows the server address and the exact values to enter in the next step.

A token gives access to the calls of a single PBX. Create one token per PBX if your account is linked to several PBXs.

#Step 2: add the connector to your assistant

Any MCP-compatible assistant can connect to Insights AI. In your assistant's settings, look for the connectors (or MCP servers) section, then add a custom connector with the following information:

Information Value
Name Insights AI
Server URL The address shown in CX-Engine, such as https://your-workspace.cx-engine.com/mcp/insights
Authentication Your access token, sent in the x-api-key header (or as a Bearer token)

Choose a connection without OAuth (often called "no sign-in" or "no authentication") and send the token in a header: OAuth sign-in is not available yet.

#Example: Claude

  1. In Claude (website or desktop app), open Customize > Connectors, then click Add custom connector
  2. Fill in the name (Insights AI) and the server address shown in CX-Engine, then click Continue
  3. Under Authentication, select No sign-in
  4. Under Request headers, choose x-api-key, paste your access token as the value and keep Required checked
  5. Click Add

Make sure to select No sign-in: with the "Sign in now" or "Sign in when needed" options, Claude attempts an OAuth sign-in, which fails.

Request headers are being rolled out by Claude: if the Request headers section does not appear, your organization does not have access yet. On Team and Enterprise plans, an owner of the Claude organization adds the connector.

#ChatGPT

ChatGPT currently only accepts connectors that use an OAuth sign-in. Connecting Insights AI from ChatGPT will be possible soon.

#Step 3: use Insights AI in a conversation

  1. Start a new conversation in your assistant
  2. Enable the Insights AI connector from the + > Connectors menu
  3. Ask your question, for example: "Which subjects caused the most frustration to our customers this month?"

Your assistant may ask for your permission the first time it uses an Insights AI tool. You can usually allow it once or for all conversations.

#Managing your tokens

From CRM Connector > Insights AI (MCP), you can at any time:

Action Effect
Disable a token Cuts access immediately, can be re-enabled later
Revoke a token Permanently deletes the token: every assistant using it loses access
Connect Shows the connection instructions again (without the token)

The Last used column tells you when a token was last used.

A token also stops working if its owner loses access to the account or the permission to use Insights AI.

Lost your token? It cannot be shown again: create a new one and revoke the old one.

#Privacy

When you connect your own assistant, the information needed to answer your questions (statistics, summaries, transcript excerpts) is sent to that assistant and processed under its publisher's terms. Make sure its use complies with your company's policy.

#For technical teams

Insights AI is a standard MCP server (Streamable HTTP transport), usable with any MCP client: Claude Code, Cursor, Claude Desktop, etc.

  • URL: https://your-workspace.cx-engine.com/mcp/insights
  • Authentication: access token in the X-API-Key header (or Authorization: Bearer)

With Claude Code:

claude mcp add --transport http insights-ai https://your-workspace.cx-engine.com/mcp/insights --header "X-API-Key: your-token"

JSON configuration (Cursor, Claude Desktop…):

{
  "mcpServers": {
    "insights-ai": {
      "type": "http",
      "url": "https://your-workspace.cx-engine.com/mcp/insights",
      "headers": { "X-API-Key": "your-token" }
    }
  }
}

All the tools listed above are read-only, except export_calls, which creates an export file for the token owner.

#Permissions

Managing tokens requires the MCP tokens permissions (crm-mcp-token.*). Administrators have them by default. Each user sees their own tokens; the crm-mcp-token.manage-any permission lets a user manage the tokens of every user of the account.

Previous The Insights AI assistant Next Export your calls