Skip to main content
Cekura tests voice agents by calling them. It plays scripted scenarios as a simulated caller, then scores the conversation against the metrics you define. AssemblyAI is a built-in provider in Cekura, so there’s no bridge to run: Cekura dials the phone number bound to your agent, and after the call it pulls that session from your AssemblyAI account.
Each Cekura result shows the transcript as AssemblyAI recorded it, not a re-transcription of the call audio. Every HTTP tool the agent called appears inline with the arguments it sent and the response it got back, so your metrics can check what the agent actually did. You need:
Cekura connects to AssemblyAI agents over the phone only. It places every call, so the agent must answer inbound calls on a number you can dial by hand.

​
1. Check the number reaches your agent

Call the number yourself before you set anything up in Cekura. If you don’t hear your agent’s greeting, Cekura’s test calls won’t reach it either. See If it does not work.
Agent ids aren’t shared between regional hosts. Phone numbers are registered on agents.us.assemblyai.com, so the agent bound to the number has to live there too. That’s the US region in Cekura, which is its default. An agent created on agents.assemblyai.com can’t answer the number.

​
2. Connect the agent in Cekura

In Cekura, go to Agents, create an agent, and select AssemblyAI as the provider. If it isn’t in the grid, click More options. Under Integration Settings, fill in: Under Telephony Settings, enter the phone number bound to the agent and leave Inbound on, so Cekura dials your number. Telephony is the only connection type for AssemblyAI, and it turns on by itself once the credentials are saved. Save the agent. The key field reads API Key Configured once it’s stored. Use Update to rotate the key later.

​
3. Run a test

Open an evaluator in Cekura, click Run, and start the run with the Telephony connection. Cekura dials your number and plays the scenario as the caller. When the call ends, Cekura finds the matching AssemblyAI session:
  1. It lists your recent sessions and keeps the ones your agent answered, from the number Cekura dialed from, that started within a few seconds of the test call.
  2. It waits for the session’s conversation record, which usually takes a few seconds after the call ends.
  3. It compares call duration and conversation content to confirm the match, so parallel test calls to the same agent each pair with the right session.
  4. It replaces its own transcript with AssemblyAI’s and attaches the session id to the result.
In the result, your agent’s turns are labeled Main Agent and the simulated caller’s turns are labeled Testing Agent. Each tool call shows up as a Function Call entry followed by a Function Call Result entry. The same call is also in your AssemblyAI account, with a stereo recording and per-turn timings. Look it up by the session id on the Cekura result. See Recordings and transcripts.

​
If it does not work

​
Configure through the API

To create or update the Cekura agent through Cekura’s API, set the provider to assemblyai and pass the credentials:
assistant_id is your AssemblyAI agent_id, and it’s required whenever a key is set. assemblyai_data.region defaults to us, so you can leave it out. The response never returns the key. Check the read-only assemblyai_api_key_configured field to see whether one is stored. For the rest of Cekura’s setup, see the Cekura AssemblyAI integration guide.

​
Cost

Every simulated call is a Voice Agent session billed to your AssemblyAI API key. Your telephony provider bills the inbound minutes, and Cekura bills its own side.