Where is NS-518206? It's been a week.
check_order
It shipped Monday with UPS and arrives Thursday.
Fine.
Passed
Automated testing for AI customer service and sales agents
Test changes to your AI agent so your customers don't have to.
Voxli plays your customers in full conversations with your agent. You see what a model switch or a prompt edit changed, and whether a new client's agent is ready to go live.
Works with any model your agent runs on
- OpenAI
- DeepSeek
- Anthropic
- Zhipu
- Mistral
- Amazon
“We had our development team spend three months building internal tools to handle what Voxli does out-of-the-box.”
Voxli's founders built ftrack and grew it from zero to millions in yearly revenue before selling it. ftrack won an Academy Award in 2021. Meet the founders
Before Voxli
A quick check misses what a change does to a real conversation.
After a model switch, a prompt edit, or a new client's setup, your agent still answers every question, and nothing looks wrong. A list of one-off test questions never gets far enough to see the difference.
Facts.
It promises a delivery date that your order system never returned.
Memory.
Five messages in, it has forgotten what the customer asked for.
Actions.
The customer only asks about a refund. It starts one anyway.
Tone.
Your brand's warm voice turns stiff and formal.
Voxli plays a real customer.
It writes the tests, runs the conversations, and shows you where one went wrong.
Connect your agent.
We set up the connection with you. No Voxli code goes inside your agent.
Build your test set.
Describe a test in plain English, or let Voxli talk to your agent and write the tests itself. You can also turn a real customer conversation into one.
Simulate your customers.
Each test is a whole conversation. The customer can be impatient, confused, or combative, and writes in their own language.
See why a test failed.
Voxli explains each failure and points to the messages that caused it. You find the problem without reading the full transcript.
Put the new version next to the old one before you ship it.
You tuned your prompts and tools to the model you use today. Another model, even a stronger one, can respond to them differently and offer a customer a discount nobody approved. Run the same tests several times on the old model and the new one, and see what got better and what got worse in each scenario.
Use cases
Agents break when something changes.
Start from the change you are about to make.
model swapprovider switch
Swap models safely
Run the old model and the new one on the same suite and see exactly what changed.
model swapRAG change
Catch hallucinations
Know when a change makes your agent invent facts, before a customer acts on one.
new client go-liveclient sign-off
Prove launch readiness
Turn the use cases, brand rules, and safety limits you agreed on into a pre-launch suite.
prompt updateknowledge base edit
Stop shipping regressions
Run the whole suite on every change, on a schedule or in CI, and get alerted when a flow breaks.
tool schema changenew tool
Trust every tool call
Check that the agent calls the right tool with the right arguments and reports what it actually did.
live issuecustomer escalation
Reproduce live issues
Import the conversation that went wrong, replay it until the failure shows itself, and keep it as a test.
We put 16 models behind one support and sales agent.
We ran every model at its vendor's default settings and tuned nothing. Use the results to pick a few models to try, then run your own tests on those.
- 1 GPT-6 Luna OpenAI 87 Quality score
- 2 DeepSeek V4.1 Flash DeepSeek 85 Quality score
- 3 GPT-5.6 Luna OpenAI 85 Quality score
For developers
Your coding agent fixes what Voxli finds
Connect Claude Code, Cursor, or any MCP client to your Voxli workspace. Your assistant reads the failing run, applies a fix, and runs the tests again.
More on MCPConnect your workspace
claude mcp add voxli --transport http https://api.voxli.io/mcpThen ask your assistant
Look up the latest run in Voxli and analyze why it failed.
- Turn a bug report into a new test
- Add checks to existing tests
- Compare two runs side by side
Bring the change you're planning.
Book 15 minutes with Mattias or Björn. We'll show you how Voxli would test that change for your agent.
Try it on your own agent, free for 14 days.
After the call we set up a workspace connected to your agent, generate a first set of scenarios from it, and hand it over. You run them, read the results, and add your own. Everything in the workspace carries over if you continue.
How pricing works