The open-source API testing tool that also drives the browser in the same plan.

It is both, and the useful framing is that it refuses to separate them. Look at assrt-mcp/src/core/agent.ts. The TOOLS array between lines 16 and 196 has seventeen tools, all registered to the same agent. Thirteen of them drive the browser (navigate, snapshot, click, type_text, select_option, scroll, press_key, wait, screenshot, evaluate, wait_for_stable, plus two email helpers). Two are assertion helpers (assert, complete_scenario). One is http_request. From the agent's point of view there is no UI/API boundary: calling http_request is the same kind of tool_use as calling click. That is the entire angle of this page. The question "is it an API tool" presumes a boundary the code does not have.

You can do that, and teams do. The cost is coordination: the Playwright test and the Postman collection are two artifacts, two runners, two report formats, and the ordering has to be maintained by CI. When the UI flow changes, the API collection and the Playwright script both need an update, and neither will fail loudly if only one is updated. With Assrt, a single #Case in plan.md owns both halves. If you delete the button but forget the API assertion, the #Case fails at the click step, not silently in a collection runner someone forgot to re-trigger.

Standard Node fetch. The handler lives at agent.ts:925-955. It takes the URL, method (default GET), headers (default Content-Type: application/json plus anything you pass), and body (only for non-GET/HEAD). It wraps the call in an AbortController with a 30-second timeout. The response body is read as text and truncated at 4000 characters with a "...(truncated)" marker, which keeps big JSON blobs from eating the LLM context window. On failure, stepStatus is set to failed and the scenario records the error, so a refused connection is a test failure, not a silent skip.

The system prompt at agent.ts:243-247 gives it the pattern directly: "When testing integrations (Telegram, Slack, GitHub, etc.): Use http_request to call external APIs (e.g. poll Telegram Bot API for messages). This lets you verify that actions in the web app produced the expected external effect." So when a #Case says "assert the Slack message arrived," the agent consults that rule, picks conversations.history, inserts the bearer token from variables, and fires the request. If you want deterministic behavior, write the URL and method into the #Case; the agent follows instructions verbatim when you provide them.

Use the --variables flag or the variables field of assrt_test. The plan references {{SLACK_BOT_TOKEN}} as a literal string; assrt interpolates the value at runtime and never logs the expanded form. The substitution happens in agent.ts inside run() before scenariosText reaches the LLM, so the raw secret is only in memory during the fetch. For CI, set the variable from a secret store. For local dev, the --extension flag lets Assrt attach to your already-logged-in Chrome, so often you do not need the token at all.

No. Each of those has a specific job Assrt does not do. Postman and Bruno are fantastic exploratory HTTP clients: you poke at an endpoint, iterate, build a request by hand. REST Assured is a library for writing Java unit tests around APIs. Karate is a DSL for contract-style assertions on REST responses. Assrt is not competing with any of them on that axis. Assrt owns a different job: the end-to-end seam where the UI triggers something and an API should reflect it. For pure spec testing of a REST endpoint in isolation, the specs-first tools are a better fit. For "the user clicked Send and the webhook was supposed to land," nothing else in the OSS landscape closes the loop in one test plan.

npx @assrt-ai/assrt setup installs the MCP server globally and wires it into Claude Code. The source is in two repos under ~/: assrt (the web UI, recorder, and dashboard) and assrt-mcp (the CLI and MCP server you actually drive tests with). Under the hood the browser is @playwright/mcp, so your tests run in real Chromium. No cloud dependency, no account required. If you want history and shareable run URLs you can sync to app.assrt.ai, but every run also writes locally to /tmp/assrt/ and the plan is just markdown.