Search Circleback transcripts for what customers said about a production bug

Last verified on Nov 8, 2026.

Customers describe bugs in calls long before they file tickets, and the details are specific: which screen, what they clicked, roughly when it stopped. Those details sit in a meeting transcript, not in your logs. Circleback records and transcribes the calls, and its MCP server lets the Sazabi agent search them. When the agent finds an error in your logs, it can check whether anyone already told you about it on a call.

What the Sazabi agent can do with Circleback

Sazabi connects to Circleback's remote MCP server at circleback.ai/api/mcp. You grant consent once through OAuth, and Sazabi refreshes the token afterward. A paid Circleback plan is required. Circleback's tool list is long, and these are the ones that matter for an investigation:

• SearchMeetings finds meetings by keyword, date range, tags, attendee, domain or folder, and ReadMeetings returns notes, attendees, action items, insights, tags, folders and duration.
• SearchTranscripts searches across all transcripts and returns matches with timestamps, and GetTranscriptsForMeetings returns full transcripts with speaker labels.
• SearchActionItems filters by keyword, status (pending or done), assignee, tag or date range.
• FindCompanies looks up a company by name or domain and returns related meetings, and FindProfiles shows a person's history across meetings.
• SearchCalendarEvents and SearchEmails cover upcoming and past events and connected email.

Circleback also has write tools such as UpdateMeeting, CreateActionItem and ShareMeeting. Sazabi governs them with the per-connection read-only toggle. On Circleback's side, the default grant is scope=user user:write, and scope=user is the read-only option.

Questions it answers

• Did any customer mention this failure on a recorded call before the first error in our logs?
• What did the person say, in their own words, and at what timestamp?
• Are there pending action items about this feature or this account?
• Which meetings have we had with this company domain?
• Who attended the review call, and what did the notes record as decisions?

A worked example: finding the call that predates the alert

Your logs show a rise in timeout errors from the /reports/generate route, and one tenant, example.com, produces most of them. You ask:

Search Circleback transcripts for mentions of "report" and "timeout" or "slow"
from calls with example.com in the last 30 days. Show me the passages with
timestamps. Then list pending action items that mention reports.

The agent runs the log query first, then calls FindCompanies with the domain to find meetings tied to that company. It follows with SearchTranscripts restricted by date and keywords, and GetTranscriptsForMeetings for the best hit so it can quote the surrounding speaker turns. SearchActionItems with status pending returns anything the team committed to.

How you use the result matters more than the lookup. A quote from three weeks ago saying exports were slow is evidence that the problem predates your alert, which changes the question from "what broke today" to "what changed three weeks ago". A quote that matches today's error text exactly is a strong lead, but it is the customer's description, so confirm it in logs. An empty result is also information, and means only that nobody recorded it. The agent can only see calls Circleback recorded.

How it fits with other integrations

Meeting records are one input among several. Granola fills the same role for teams that record calls there, Attio holds the account and owner, and Linear is where an action item becomes tracked work. See all connectors under integrations and plans on pricing.