Cekura

What is Cekura?

Cekura is a platform for automated testing, evaluation, and production monitoring of voice and chat AI agents; it is operated by the American company Tatva Labs Inc. Teams can simulate real conversations before the systems go live, monitor calls continuously after they are launched, and convert any issues identified into repeatable regression tests.

It primarily serves the product, engineering, and quality teams responsible for developing voice-based customer service systems, telemarketing solutions, booking systems, recruitment tools, and compliance-related dialogue systems. The focus of these efforts is not on creating new voice robots, but rather on determining whether existing robots can function reliably in different scenarios, on various platforms, and with different infrastructure setups.

A one-sentence summary

Cekura enables bulk calls to AI agents using synthetic users, allows scoring based on custom metrics, and monitors real-world conversations; this facilitates the team in identifying issues such as regressions, delays, tool calls, and compliance problems using the same evaluation criteria before and after deployment.

Main functions

Dialogue simulation before going live

Users can create test scenarios for tasks such as reservations, refunds, identity verification, or determination of sales eligibility, and then have simulated users engage in multiple rounds of conversation with the target intelligent agent. Tests can include background noise, accents, multiple languages, interruptions, and complex branching scenarios, rather than merely comparing a fixed text.

Evaluators and test subjects

The evaluator is used to define the identity, objectives, speaking style, conditional actions, and expected outcomes of simulated users. Test data, dynamic variables, and responses from simulation tools can be reused, allowing the same scenario to apply to different customers, locations, and boundary conditions.

Predefined and custom metrics

The platform provides standard metrics such as accuracy, conversation quality, customer experience, and voice quality, and it supports boolean, numerical, and categorical results. Teams can also create LLM evaluation metrics using natural language, or write custom deterministic checks in Python.

Tool invocation and infrastructure testing

Cekura can verify whether an agent calls the correct tools at the right time and passes the proper parameters; it can also use simulation tools to produce stable results. Infrastructure testing covers various issues related to voice connections, such as latency, stability, disconnections, silence, interruptions, concurrency, and recovery from failures.

A/B, load, and multilingual testing

A team can use the same scenario to compare two prompts, models, suppliers, or configurations, and observe uncertainty by running tests repeatedly. Load testing and concurrent calls are used to assess capacity, while tests involving multiple languages, accents, and code switching are intended to reveal defects in recognition and understanding.

Production call monitoring

The production system can send transcripts, recordings, metadata, and reasons for completion either through native integration or via Webhooks. Cekura evaluates calls based on various metrics, displaying trends, failure points, topics, delays, and anomalies, and it also provides alerts and downloadable reports.

Red Team and Compliance Assessment

The Red Team feature enables the creation of adversarial conversations such as jailbreaking, deviation from scripts, malicious intentions, and leakage of personal information. Medical, financial, or other regulated industries can also implement rules such as authentication, audio recording notifications, and restrictions on information disclosure as part of reusable access controls.

Self-improvement feedback loop

The platform generates or recreates tests based on production failures, clusters the underlying causes, and suggests prompts for corrections, after which it runs the release process again. The automatic suggestions still need to be reviewed by engineers, especially when changes relate to business policies, security, and compliance rules.

Closed loop of testing, monitoring, and improvement

Complete workflow

1. Create organizations and projects, and categorize agents and access permissions by environment, product line, or customer.
2. Connect to Retell, Vapi, ElevenLabs, LiveKit, Pipecat, SIP, or custom voice systems.
3. Select key business processes to create scenarios for normal operations, edge cases, failure recovery, and compliance risks.
4. Set up simulated characters, accents, languages, dynamic variables, tool responses, and call termination conditions.
5. Select predefined metrics, and use an LLM to determine the scores or add business-specific metrics and scoring rules via Python.
6. Run small samples with low concurrency, and manually review the transcriptions, recordings, explanations of metrics, and any errors.
7. Expand to regression, A/B, load, and red team testing, and integrate key tests into pull requests or deployment pipelines.
8. Integrate with production call monitoring, and configure alerts, personal data handling, log retention, and access control.
9. Convert frequent failures into new test cases, revise the prompts or infrastructure, and then verify again.

Tutorial for beginners

1. After registering for free and creating a new project, connect to a voice or chat AI in a non-production environment first.
2. Start with the most important part of the user journey, and clearly define the goals of the simulated user, the information available to them, and the conditions that need to be met.
3. Create three to five scenarios, covering normal completion, user interruption, missing information, and tool failure.
4. Add metrics such as task completion, accuracy of facts, latency, voice quality, and compliance.
5. Run a small number of tests and listen to or read the results one by one to check whether the scores match the manual judgments.
6. Adjust the vague indicators and simulated characters, and add accents, noise, language, and boundary conditions.
7. Once testing is stable, create smoke and full regression test suites based on the tags, and establish deployment controls.

Integrate with the CI/CD process

1. Create an API key in Cekura and save it in the encrypted secrets of the code hosting platform.
2. Record the ID of the target agent, and select the test set to be executed using the scenario ID or tags.
3. In a GitHub Actions workflow, call the official Action and set the execution frequency, timeout, and optional phone number.
4. Enable the workflow to run when a pull request is submitted, a change is pushed to the main branch, a scheduled task occurs, or it is triggered manually.
5. Review the operation logs and test results; subsequent deployment is permitted only when the key metrics reach the specified thresholds.
6. Rotate keys regularly, and avoid storing phone numbers, customer data, and credentials directly in the repository.

Which users are it suitable for

• Voice AI development team: Runs automatic regression tests for every change to prompts, models, voice stacks, or tools.
• Quality and Testing Engineers: Convert dialogue business requirements into scenarios, metrics, and release criteria.
• Product and Operations Team: Identifies failed processes, disconnections, and customer experience issues from recorded calls.
• Customer service and outbound calling platform: Testing of booking, refunds, eligibility assessment, call forwarding, and pathways for human intervention.
• Healthcare and regulated industries: identity verification, audio recording for notification, sensitive information, and business compliance rules.
• Voice platform providers: Compare models, orchestration layers, and communication infrastructure in the same scenario.
• AI consulting and implementation team: Maintains separate projects, reports, and repeat testing assets for different clients.

Typical use cases

• Simulate hundreds of scenarios for reservations, date changes, cancellations, and emergency upgrades before release.
• Compare the performance of Vapi, Retell, ElevenLabs, LiveKit, or a custom-built voice stack in the same scenario.
• Test whether the agent can correctly use calendar, CRM, payment, or knowledge base tools.
• Test the robustness of voice using accent, background noise, interruptions, long silences, and network disruptions.
• Replay actual production failures as regression tests to prevent their recurrence after fixes are applied.
• Monitor a large number of actual calls and cluster them by the reasons for failure, thereby reducing the need for manual listening to each call.
• Run smoke tests in GitHub pull requests to prevent the deployment of versions that could cause critical workflows to fail.

Product advantages

• By including both pre-launch simulations and post-launch monitoring within the same set of metrics, it becomes easier to go back to repeatable testing in the event of production issues.
• It supports both LLM-based evaluation and Python metrics, taking into account both semantic judgment and deterministic business checks.
• It covers voice details, tool invocation, infrastructure, load, red teaming, and compliance, and is not limited to text-based Q&A.
• It offers native integration with various mainstream voice platforms, and also supports SIP, WebSocket, Webhook, as well as custom connections.
• Public REST APIs, OpenAPI specifications, Python SDKs, CLI tools, MCP, and CI/CD tools are available, providing a complete set of automation options.
• The usage plan allows one seat for free, which is suitable for creating a small test set using real agents first.

Usage restrictions and precautions

• Simulated users and LLM judges themselves are also subject to uncertainty; a successful test run once does not guarantee that the system will be reliable in a production environment.
• Unclear indicator design can lead to incorrect judgments; therefore, the team needs to use manually labeled samples to calibrate the scoring rules and thresholds.
• Voice test duration, chat responses, monitored calls, seats, and concurrency each affect the cost, while load testing can quickly deplete the quota.
• Different platforms use various formats for transcription, recording, tool invocation, and timestamps; incomplete integration can limit the level of observability.
• Production calls may contain personal, health, payment, or customer confidential information; therefore, it is necessary to set up mechanisms for data masking, access controls, retention policies, and contractual terms.
• The quota under the Startup plan is based on shared points; it amounts to around 2,000 minutes for testing and around 10,000 monitoring sessions. These are not two separate unlimited quotas that are granted simultaneously.
• The data on accuracy, precision, and customer efficiency provided on the official website represent the results of the supplier or of specific samples, and cannot be directly applied to one’s own intelligent agent.
• The automatically generated prompt patches may alter the security boundaries or business logic, and therefore must undergo code review and regression testing.

Prices and packages

As of August 21, 2026, Cekura offers three subscription plans: Usage-based, Startup, and Enterprise. New accounts receive 300 free credits, which is equivalent to approximately 60 minutes of voice testing on the official website; no credit card is required to get started.

Startup offers monthly subscriptions that can be canceled at any time, while Enterprise relies on annual contracts and supports payment via purchase orders, invoices, or bank transfers. The rates, taxes, point conversions, and final terms are subject to the account settlement page and the contract details.

Key differences among the various options

Free quota and billing method

Each plan includes features such as call simulation in production environments, alerts, downloadable reports, standard metrics, Python-based metrics, APIs, MCP, and an AI coding assistant. 300 free credits are provided to get started, but additional seats as well as usage for actual testing or monitoring incur costs.

When estimating usage, it is necessary to calculate separately the minutes spent on voice tests, the number of chat responses, the count of monitored calls, the peak level of concurrent connections, the frequency of re-running tests, and the amount of time required for log retention. If a test suite is run multiple times each time it is submitted, the actual monthly consumption will be significantly higher than that of a single manual test.

Supported agents and integrations

API, Python SDK, and MCP

GitHub and open source

Cekura’s official GitHub repository includes an integration for GitHub Actions as well as a Skills repository designed for AI coding assistants; both are licensed under the MIT license. The Skills repository covers various workflows such as creating agents, designing evaluations, generating scenarios, fixing issues in production environments, and building infrastructure tests.

These repositories only provide integration tools, workflows, and documentation; they do not contain the full source code of the Cekura hosting testing platform. The platform itself remains a closed-source commercial service, and the fact that the Python SDK can call public APIs does not mean that the core services are open source.

Security, privacy, and data processing

The official website highlights the security and compliance capabilities related to SOC 2, HIPAA, and GDPR. The Startup package includes the signing of BAA and DPA agreements, while the Enterprise package can be customized. When dealing with protected health information, it is necessary to ensure that the relevant agreements have been signed and that the functions being used fall within the scope of coverage before uploading any actual calls.

Production monitoring deals with transcription, recordings, phone numbers or identifiers, tool calls, and business metadata – content that may contain sensitive information. While documents provide the capability to mask personal data, it is still up to the team to decide which fields should be removed before they are sent.

Sub-processor

The public sub-processor page lists Supabase for data storage and database management, as well as OpenAI, Anthropic, and Replicate for model capabilities. Companies should determine which services will be used in their actual projects, the location of the data, the retention period, and the procedures for deleting it once use is ceased.

Retention, Accounts & Subscriptions

The retention period for package logs is either 30 days, 90 days, or can be customized according to the company’s requirements. The privacy policy dictates that account personal data should be retained only for as long as it is necessary to achieve the intended purposes. The terms state that subscriptions are automatically renewed, and cancellation takes effect at the end of the current payment period; prices may also be adjusted after appropriate notification in accordance with the law.

List of security practices

• Prioritize the use of synthetic identities, test numbers, and fictional sensitive information to create pre-release scenarios.
• Configure PII masking before connecting to production, and send only the transcription, recordings, and metadata required for evaluation.
• Different API keys are used for development, CI, and production, with permissions, IP addresses, and project scopes being restricted.
• Store the keys in the Secrets management tool; do not include them in the code, logs, or test reports.
• Select the log retention period in accordance with regulations and business requirements, and establish procedures for deletion and handling requests from data subjects.
• For medical projects, a BAA is required; for projects in Europe or other regions, a DPA, as well as arrangements regarding cross-border data transfer and data storage, are necessary.
• Regular audits, alerts, and manual inspections are employed to prevent the testing platform from becoming a new hub for sensitive data.

Basic information

Recommendation score

The recommendation score is 4.6 out of 5. Cekura integrates scenario simulation, semantic and code metrics, voice infrastructure testing, production monitoring, and CI/CD into a single lifecycle framework, which is highly valuable for teams that are ready to deploy voice agents or those that are continuously working on their improvement.

The main challenges are that high concurrency and frequent updates can quickly lead to high costs related to usage, while obtaining reliable results relies on the team’s ability to define clear metrics and carry out manual adjustments. When dealing with recorded audio from production environments, privacy, contractual matters, and the management of sub-processors cannot be left to the platform’s default settings.

Frequently Asked Questions

Can Cekura be used for free?

New accounts come with 300 free credits, which are equivalent to approximately 60 minutes of voice testing on the official website; the first test session is free and no credit card is required. Subsequent tests are charged based on the number of minutes used, the number of replies sent in chats, or monitored calls, and it is also possible to upgrade to a monthly subscription plan.

Can it only test voice agents?

Its core function is that of a voice intelligent agent, but it also supports cost-effective logical testing via chat simulations, text messages, WhatsApp, and plain text. The audio quality, latency, and interruption levels associated with each communication channel vary.

Is it possible to monitor actual production calls?

Yes. Native integration or Webhooks can be used to send transcripts, recordings, and metadata, which are then used by metrics evaluation, dashboards, and alerts to help teams identify any abnormalities.

Which voice platforms are supported?

The official documentation lists Retell, Vapi, ElevenLabs, LiveKit, Pipecat, Bland AI, SIP, Agora, as well as custom integrations. The specific testing and monitoring capabilities should be checked according to the guidelines provided for each platform.

Are APIs and SDKs provided?

It offers public REST APIs, OpenAPI specifications, a Python SDK and CLI, as well as MCP servers and GitHub Actions. API calls are subject to restrictions related to account permissions, credits, concurrency, and plans.

Is Cekura open source?

The core hosting platform is not open source. While GitHub Action and the AI coding assistant Skills are available under the MIT license, this does not allow the entire product to be labeled as open source.

Are the LLM evaluation results reliable?

LLM judges are suitable for assessing semantics and the quality of conversations, but they may make mistakes or change over time as the models evolve. Important metrics should be calibrated using manually labeled samples, and rules that allow for deterministic calculations should preferably make use of Python-based metrics.

Can the Startup plan allow for testing 2000 minutes at the same time and monitoring 10,000 messages?

It cannot be understood in such a simple way. The two figures represent approximate conversions by the official website for the monthly credit limit; when used together, they both result in a consumption of credits.

Can it be deployed in a local or private environment?

The enterprise plan offers VPC and on-premises deployment options; a customized contract and technical assessment are required. Managed services are available based on usage or through the Startup plan.

What should be taken into account regarding production data?

First, verify the legal basis for recording and processing; thereafter, impose restrictions on transcription, recording, as well as on the transmission of numbers and health-related information. It is also necessary to establish procedures for data masking, retention, permission management, as well as DPA or BAA agreements, along with processes for data deletion.

Summary

Cekura offers a comprehensive quality management system for voice agents, covering everything from stress testing prior to deployment to replaying failures that occur in production. It is particularly suitable for teams that already have real business processes, need to release updates frequently, and wish to integrate evaluation processes into their development workflow.

The tool itself does not guarantee reliability automatically; true quality comes from clear scenarios, interpretable metrics, manual calibration, and strict data governance. It is usually more manageable to first establish a baseline using a small set of tests related to critical paths, and then gradually expand parallel processing and production monitoring.