AI software testing: who owns the tests after you cancel?

Test cases are stored as plain markdown using a '#Case N: name' heading format. Each case is a numbered list of natural language instructions: navigate here, fill this field, assert that text. Assrt writes the live plan to /tmp/assrt/scenario.md and caches completed scenarios as JSON at ~/.assrt/scenarios/<uuid>.json. The JSON wraps the same markdown string, so the test plan is always readable in a text editor and committable to git.

With most commercial platforms, your tests are stored in their proprietary format inside their cloud. The access token that lets you fetch them is tied to your paid subscription. Cancel and the token is revoked; you get at most a CSV export of test names if the vendor is generous. With Assrt, the UUID that identifies a scenario also functions as the access token — no subscription required. Tests are cached locally before and after each run, so cancelling an API plan (there is none) or losing connectivity does not erase anything.

Each Assrt scenario is identified by a UUID v4. That UUID IS the access token: anyone who knows the UUID can load the scenario from app.assrt.ai without presenting any separate credential. This is the capability-URL model, the same pattern used by shared Google Docs with 'anyone with the link' access. Practically: sharing a test means sharing a UUID. Revocation means discarding the UUID. The source comment at scenario-store.ts line 8 spells this out: 'No auth required: the UUID v4 IS the access token.'

Yes. After any scenario runs, the full scenario object (plan, URL, variables, tags) is written to ~/.assrt/scenarios/<uuid>.json. The next run checks the local cache first; it only calls app.assrt.ai to sync updates or load a scenario you have not seen before. The MCP server itself runs on your machine or your CI, so the execution path is entirely local: Claude Haiku drives the browser through Playwright MCP, results write to /tmp/assrt/results/, and the video player is a localhost HTTP server. Nothing in the test loop requires an external service.

AI-generated test code produces a .spec.ts or .yaml file that a conventional runner then executes. The AI is a one-time author; all adaptation to UI changes happens at the next authoring session. LLM-driven execution never produces a script file. The model receives a live accessibility tree snapshot, sees element references like ref=e5, and decides each tool call turn-by-turn. When the UI changes between two runs, the model binds to whatever element name matches today, not the name from yesterday's authoring session. The adaptation happens at click time, not at the next code-gen cycle.

MCP (Model Context Protocol) is an open standard for connecting AI models to tool surfaces. Assrt's test runner is implemented as an MCP server, which means the tool definitions are expressed in the same schema that any MCP-compatible model can consume. You can run Assrt from Claude Code, from any IDE with MCP support, or from the CLI. Because the protocol is open, the runner is not coupled to a single AI vendor: swap the model parameter and the same scenario runs against a different underlying LLM. Proprietary APIs for AI testing tend to lock both the format and the AI layer simultaneously.

The simplest approach: commit the test plan markdown files alongside your application code. Assrt watches /tmp/assrt/scenario.md with a 1-second debounce and syncs edits to the cloud; you can also commit that file directly or maintain it in a tests/ directory. The local cache at ~/.assrt/scenarios/<uuid>.json is also plain JSON — git-committable, diff-readable, mergeable. A team workflow looks like: write test cases in markdown, commit them, run assrt run on CI pointing at the committed file, push results as CI artifacts.

Assrt MCP is open-source and free to install and run. Each test run uses Claude Haiku via the Anthropic API, which charges per token. A typical 10-step scenario runs to about 20,000 input tokens and 2,000 output tokens per turn, across roughly 10 turns, putting the per-run API cost in the range of a few cents depending on Anthropic's current pricing. There is no Assrt subscription, no per-seat fee, and no vendor-locked cloud where your tests live. The cost ceiling is your Anthropic API usage.