NewCompare models in chat agent tests.See the benchmark

Test your chat agent before launch and after every change.

Voxli simulates users in full conversations, checking your agent's answers and actions against your rules.

The Voxli run results page

Let Voxli AI write the tests.

Describe what to test. Voxli AI explores your agent, writes tests, and runs them.

Edit the plain-English instructions and checks, then reuse the tests after your next change.

Learns how your agent works

Voxli AI chats with your agent to learn its workflows and rules before it writes tests.

Voxli AI: I want to return my headphones.

NorthSphere Agent: Sure. What's your order number?

Knows where agents often fail

Voxli AI adapts known failure patterns to your agent: ignored corrections, skipped approvals, and unfinished tasks.

Voxli AI: Cancel both of my orders.

NorthSphere Agent: Done. Your first order is canceled. Canceled only one order

Run full conversations and check what your agent did.

Same task, different users

Use Personalities to test the same task with different tones and languages.

  • Neutral: Sorry, I meant my headphones.
  • Impatient: No, not the speaker. The headphones.
  • French speaker: Pardon, je voulais dire mon casque.

Catch failures that don't happen every time.

Repeat each test in separate conversations to see whether your agent handles the task consistently.

  • Attempt 1 Passed
  • Attempt 2 Passed
  • Attempt 3 Failed

Check the action behind the reply

Use recorded tool calls to check what your agent actually did.

Customer

Sorry, I meant my headphones.

Agent

Done. I've started your headphones return.

start_return Check failed

Item: Speaker

Expected headphones

Hallucination detection.

Find claims your agent can't support.

After a test, check factual claims against the recorded conversation and tool results.

Recorded order lookup

Status
Shipped
Delivery date
Not available

Agent reply

Your order has shipped.

It will arrive tomorrow.

Unsupported claim

The order lookup gave no delivery date.

Compare versions. See what improved and what broke.

Compare runs the same tests on your current agent and a proposed version with a different model, prompt, knowledge base, or tool.

Ask Voxli AI what your results mean.

Voxli AI summarizes a run, a comparison, or a single test, finds patterns across conversations, and links to the evidence.

Voxli AI

How does the new model compare with the current one?

It's cheaper and faster, but it scores lower and hallucinates more.

  • Costs a third as much and replies 1.5 seconds faster.
  • Product questions: up from 92% to 98%.
  • Billing and refunds: down from 96% to 72%. It often ignores customer corrections.
  • German speakers: down from 93% to 71%. It often replies in English.

Fix what Voxli finds with your coding assistant.

Bring failed conversations and checks into Claude Code or Cursor through MCP. Ask your assistant to fix the code or prompt, then rerun the tests.

Connect your coding assistant

Tests can also run in CI with GitHub Actions.

Connect your workspace

claude mcp add voxli --transport http https://api.voxli.io/mcp

Then ask your assistant

Investigate the latest failed run, help me fix the agent, and rerun the tests.

  • Read conversations message by message
  • Inspect failed checks and reported tool calls
  • Rerun tests and compare runs
“Engineers were putting in the same question again and again, hoping the issue would show up. You'd ship a fix, then run it 10 times hoping it doesn't come back. It was absolute insanity.”
Ammar Khan Read the Expertise.ai story

Find problems in real chats and test for them.

Import your agent's real chats to see what your tests may be missing.

Inspect recorded tool calls from Langfuse, LangSmith, or OpenTelemetry traces.

  1. Real conversations

    Find issues with flags and labels.

  2. Voxli AI investigates

    Try to reproduce the issue.

  3. Reusable regression test

    Test for it after future changes.

    When reproduced

Keep monitoring your agent after release.

Run key tests on any recurring schedule, such as every hour or each weekday at 9 AM. Follow scores over time as models, tools, and knowledge change.

Choose which failures trigger Slack or email alerts.

Nightly support tests Blocker failed

Test score68%

Connect the agent you want to test.

Adapt our sample script to call your agent, then run tests locally or in GitHub Actions.

A founder can help you connect your agent. Book a call to work through your setup together.

Mattias LagergrenBjörn Rydahl

Run your first conversation test.

Try Voxli free for 14 days. Book a call if you want help getting started.

Start free trial Book a call