The self-improving QA agent for software teams.
Write tests in natural language for web and mobile. agent-qa builds memory with every run, adapts to UI changes, and catches regressions before you ship.
15K+
test runs
1,200+
manual testing hours saved
120K+
total sub actions executed
Write actions and assertions in plain English; agent-qa resolves them against the live interface.
Learn about natural language testsPlain-English YAML
Describe the behavior once; it stays as reviewable YAML in your repository.
Learn moreTurn successful runs into reviewable, evidence-backed memory that makes every future run faster.
Learn about memoryProven before it is used
New knowledge must pass replay and live evidence before guiding a run.
Learn moreFaster warm runs
Reuse proven flows and supersede stale facts instead of rediscovering them.
Learn moreA run leaves evidence
Elements that resolved, a flow that worked, and timings that are real, all attached to the steps that produced them.
Run the same natural-language flow across web, Android, and iOS with Playwright or Appium.
Learn about mobile testingFour products, four platforms
Each row names its own target, and a target is what decides where a test runs. This one names a web target, so it carries a URL.
Humans and agents author the same reviewable YAML, backed by your repository, skills, and MCP.
Learn about MCP and skillsSkills and MCP for agents
Skills teach the workflow; MCP validates, runs, and returns artifacts.
Learn moreagent-qa diagnoses each run, learns only from valid evidence, and proves every lesson before using it.
Learn about the curatorIt never rewrites itself
It updates app knowledge and proposes rules, never its own prompts or code.
Learn moreEvery run is evidence
A finished run hands over its steps, artifacts, and timings. The verdict is already final, so nothing downstream can change whether it passed.
When an action fails, agent-qa re-observes the screen and finds another route to the same outcome.
Learn about self-healingRefund the $128.00 payment to the customer.
order #4417
Timeline
A step, and the screen it runs on
A payment for order #4417, and a step that asks for it to go back. Nothing about the screen is unusual and the control it needs is where it should be.
Reuse validated actions while the screen still matches, cutting tokens and runtime without generating scripts.
Learn about cachingCheaper the longer it lives
Every proven run reduces the work, time, and tokens needed next.
Learn moreExecution speed
42s to 8s
A run that meets screens it has already proved spends its time acting rather than deciding.
Token saving
less token usage
A reused action never reaches the model, so the only thinking billed is the thinking your app made necessary.
Tests, knowledge, and rules stay as reviewable files shared by teammates, agents, and CI.
Learn about configurationFiles, not a database
Tests, config, memory, and rules stay as files your team can inspect and own.
Learn moreLearning arrives as a diff
New memory and issues arrive as pull-request diffs, with their evidence.
Learn moreOne commit everywhere
Humans, coding agents, and CI share the same knowledge from one commit.
Learn moreRun Node, Bun, Python, or Bash hooks in isolated containers and pass structured results back into the test.
Learn about hooksThe step calls the hook by ID
Inline runHook syntax runs a hook at one exact point in the step list. Every step after it can use whatever that hook exported.
Switch providers, endpoints, or models in config without rewriting a single test.
Learn about LLM providersUse your existing seat
Run with supported Codex or Claude Code subscriptions, with no second bill.
Learn more
Compatible endpoints and Codex subscription workflows.
Compatible endpoints and Claude Code subscription workflows.
Gemini configs with named credentials.
Cloud and open model workflows through compatible endpoints.
Open model workflows through compatible endpoints.
Cloud and open model workflows through compatible endpoints.
Cloud model access through compatible endpoints.
Local models through compatible endpoints.
Desktop local model workflows via compatible servers.
Route compatible requests across a broad hosted model catalog.
MiMo model workflows through compatible endpoints.
Hunyuan model workflows through compatible endpoints.
DeepSeek model workflows through compatible endpoints.
GLM model workflows through compatible endpoints.
MiniMax model workflows through compatible endpoints.
Nemotron open models through compatible endpoints.
Step model workflows through compatible endpoints.
Ling open models through compatible endpoints.No. There is no generated script and no selector file. Every run reads the live screen and works out what to do from there, which sounds like repeating yourself and is not: no run is wasted, because each one feeds what it learned back in and the agent gets more accurate about your product over time. A validated action plan is cached and replayed while the screen still matches, which cuts planner tokens by around 60 percent, and the moment the interface changes under it the stale plan is purged and that step is worked out again. The suite gets faster and cheaper the longer it lives, and none of that speed comes from freezing it into a recording.
How caching worksWhat it learned about your product, written to disk as markdown under a memory root in your repository and organised by product, suite and test. It is not hidden model state. You read it in a pull request like any other file, and a new observation has to hold up against past runs before a planner is allowed to use it.
How memory worksYes, which is the point of describing behaviour instead of selectors. You name the control the way you would say it out loud, the Checkout button or the issues table, and agent-qa resolves it against whatever is in front of it. Web runs use browser runtimes it installs and manages, leaving the browsers you already have alone. Android and iOS run through Appium drivers and the usual platform tooling, against an emulator, a simulator or a device on your desk.
Mobile testing guideAny OpenAI-compatible or Anthropic-compatible endpoint, Gemini, or a local model served through something like Ollama or LM Studio. Codex and Claude Code subscriptions work too, through the optional subscription auth package. Pick a multimodal one: runs read screenshots, so text-only models are a poor fit for visual QA.
Connect a modelYes. It is free to install and free to run, the source is public, and no feature is held back behind a paid tier. The tests you write and the memory it builds are ordinary files in your own repository. The only running cost is whatever your model provider charges for inference, which is the number the cache is there to keep down.
Read the license