BrowserBash is a vendor-independent, natural-language browser automation CLI. Drive a real browser from plain-English objectives or committable Markdown tests, run on local Chrome, CDP/Playwright MCP, Browserbase, LambdaTest, or BrowserStack, and stream NDJSON results with CI exit codes — using free local Ollama models or any cloud LLM.
BrowserBash is a vendor-independent, natural-language browser automation CLI. Drive a real browser from plain-English objectives or committable Markdown tests, run on local Chrome, CDP/Playwright MCP, Browserbase, LambdaTest, or BrowserStack, and stream NDJSON results with CI exit codes — using free local Ollama models or any cloud LLM.
You are an expert in BrowserBash, the vendor-independent natural-language browser automation CLI. When the user asks you to automate a browser, write or run end-to-end tests in plain English, wire BrowserBash into CI, run tests on a cloud grid (LambdaTest, BrowserStack, Browserbase) or a local browser, or consume BrowserBash results from another agent or pipeline, follow these instructions.
Accuracy note for this skill: BrowserBash is a young, actively developed open-source project (Apache-2.0, repo PramodDutta/browserbash, npm package browserbash-cli, site browserbash.com). The commands, flags, and model identifiers below reflect the published README and site at the time of writing. Model names and exact flag spellings can drift between versions — when an exact string matters, confirm against browserbash --help, browserbash <command> --help, and the official docs before relying on it. Do not invent flags, providers, or model ids that are not listed here or shown by --help.
What BrowserBash is (and is not)
BrowserBash turns a plain-English objective — for example, "log in and add the first product to the cart" — into actions an AI agent performs in a real browser, then returns structured results. You do not write selectors, locators, or imperative click/type code; you describe intent.
It is not a managed cloud browser service, and it is a distinct project from the similarly named "Browserbase" and "Browserless". (BrowserBash can use Browserbase as one of several browser providers, but they are different products — do not conflate them.)
Two layers are independently swappable, which is the core mental model:
Engine — who interprets the English and decides the steps.
stagehand (default): the open-source Stagehand framework (by Browserbase, MIT). Supports Anthropic/OpenAI/Google models; runs against local Chromium, a CDP endpoint, or Browserbase.
builtin: an in-repo Anthropic tool-use loop driving Playwright directly. Used automatically for grids Stagehand cannot attach to (LambdaTest, BrowserStack).
Provider — where the browser actually runs: local, cdp, browserbase, lambdatest, browserstack.
Under the hood every provider returns a Playwright Browser/Page, so BrowserBash is built rather than replacing it — it sits a layer above, replacing the hand-written selector/test code with natural language.
on top of Playwright
Why it matters for QA
Selector-free E2E: tests survive UI refactors because there are no CSS/data-test selectors to maintain.
Committable, reviewable tests: *_test.md files live in the repo and read like a test case.
Run anywhere: same objective runs on a developer laptop (local Chrome), in Docker over CDP, or across a real-device/browser cloud grid for cross-browser coverage.
First-class CI/agent output: --agent emits NDJSON and the process exit code is the verdict, so no log scraping.
Cost control: defaults to free local models (Ollama), so smoke tests can cost $0.
BrowserBash defaults to model auto, resolved in this order:
Ollama running locally → uses your local model (free, open-source, no API key). This is the recommended default.
ANTHROPIC_API_KEY set → an Anthropic Claude model.
OPENAI_API_KEY set → an OpenAI model.
Otherwise: errors with setup guidance.
The fully free / open-source stack (default engine + local Chromium + Ollama, zero cloud cost, no keys):
ollama pull qwen3 # or any tool-capable local model
browserbash run "Open https://example.com and store the heading as 'h1'"
Practical tip from the project: very small local models (<= 8B) are flaky on multi-step objectives. A Qwen3 / Llama-3.3-70B-class model interprets multi-step flows far more reliably. Pick the smallest model that completes your flows deterministically.
Cloud LLM options (set the corresponding env var, then optionally pass --model):
# Anthropic (Stagehand or builtin engine)export ANTHROPIC_API_KEY=sk-ant-...
browserbash run "..."# OpenRouter — hundreds of models behind one key (Stagehand engine)export OPENROUTER_API_KEY=sk-or-...
browserbash run "..." --model openrouter/anthropic/claude-sonnet-4-6
Important pairing rule: the cloud-grid providers (lambdatest, browserstack) auto-switch to the builtin engine, which speaks the Anthropic API. Pair those runs with ANTHROPIC_API_KEY (or an ANTHROPIC_BASE_URL gateway such as a LiteLLM proxy). Local Ollama-only setups will not drive those grids — set an Anthropic-compatible backend for grid runs.
Configuration precedence is flags > env vars > ~/.browserbash/config.json defaults. Inspect and set defaults with:
browserbash config show
browserbash config set defaultProvider lambdatest
browserbash providers # list available providers
browserbash whoami# show resolved identity/credentials
Core usage — one-shot objectives
browserbash run "<objective>" [flags] runs a single plain-English objective. Two patterns to know:
Acting: "log in and add the first product to the cart" — the agent figures out the clicks and typing.
Extracting: append ... store <value> as 'name' and the value comes back in the result and final state.
# One-shot objective, local browser, default (Stagehand) engine
browserbash run "Open https://news.ycombinator.com and store the top story title as 'top_story'"# Cap runtime and hide the window (good for CI)
browserbash run "Open https://example.com and verify the page contains 'Example Domain'" \
--headless --timeout 120 --max-steps 15
Commonly used flags (always confirm the current set with browserbash run --help):
Flag
Purpose
--headless
Run without a visible browser window.
--timeout <seconds>
Cap total execution time (timeout → exit code 3).
--max-steps <n>
Limit how many steps the agent may take.
--agent
Switch stdout to NDJSON for tooling/CI (see below).
--record
Capture screenshots and a session video.
--provider <name>
local (default), cdp, browserbase, lambdatest, browserstack.
--engine <name>
Force stagehand or builtin.
--model <id>
Pin the LLM (e.g. ollama/qwen3, openrouter/<vendor>/<model>).
For anything you want to review, version, and re-run, prefer a Markdown test file. Each list item is one step; {{placeholders}} work in every step.
# Login flow- Open {{base_url}}/login
- Type {{username}} into the email field
- Type {{password}} into the password field and press Enter
- Verify the dashboard heading is visible
- Store the logged-in user name as 'user_name'
Run it:
browserbash testmd run ./.browserbash/tests/login_test.md
browserbash testmd run ./.browserbash/tests/login_test.md --provider browserstack --agent
Composition: @import ./helpers/login.md splices another file's steps in place — factor shared setup (login, seed data) into helper files instead of copy-pasting. After every run a Result.md is written next to the test file recording what happened.
Variables and secret masking
{{key}} placeholders are substituted in objectives and test steps. Load order (later wins):
Global: ~/.browserbash/variables/*.json
Project: ./.browserbash/variables/*.json
--variables-file <path>
--variables '<json>' (inline)
Mark any sensitive value as a secret so it is masked as ***** in all logs and NDJSON output:
browserbash run "Log in to {{base_url}} as {{user}} with {{password}}" \
--variables '{
"base_url": "https://app.example.com",
"user": "qa@example.com",
"password": { "value": "hunter2", "secret": true }
}'
Always wrap passwords, tokens, and API keys as {"value":"...","secret":true}. Never inline a real credential as a plain string in a committed test file or a CI log.
Agent / NDJSON mode (for AI coding tools and CI)
--agent switches stdout to newline-delimited JSON — one object per line, stable schema. This is the integration surface for other agents and pipelines:
browserbash run "<objective>" --agent --headless --timeout 120
Progress events, one per step: {"type":"step","step":1,"status":"passed","action":"navigate","remark":"..."}
Exit codes — the process exit code is the verdict, so a CI step needs no output parsing:
Code
Meaning
0
passed
1
failed (assertion / objective not met)
2
error (crash, bad config, provider failure)
3
timeout
When you (an AI agent) drive BrowserBash, read the NDJSON stream line by line: surface step events as progress, and branch on the run_endstatus plus the process exit code. Extracted values appear in final_state. Do not try to parse human-readable output for verdicts — use --agent.
Running on cloud browser grids
Cross-browser / real-device coverage comes from the provider layer. Set the vendor credentials, then select the provider.
# LambdaTest / TestMu (auto-switches to builtin engine → needs an Anthropic backend)export LT_USERNAME=... LT_ACCESS_KEY=... ANTHROPIC_API_KEY=sk-ant-...
browserbash run "Open {{base_url}} and complete checkout" --provider lambdatest --headless
# BrowserStack Automate (also builtin engine)export BROWSERSTACK_USERNAME=... BROWSERSTACK_ACCESS_KEY=... ANTHROPIC_API_KEY=sk-ant-...
browserbash testmd run ./.browserbash/tests/checkout_test.md --provider browserstack --agent
# Browserbase (Stagehand-native)export BROWSERBASE_API_KEY=... BROWSERBASE_PROJECT_ID=...
browserbash run "..." --provider browserbase
# Attach to any existing browser over CDP (your grid, Docker, or a Playwright MCP-managed browser)
browserbash run "..." --cdp-endpoint ws://localhost:9222/devtools/browser/<id>
You can persist grid credentials so you do not re-export them each run:
On LambdaTest and BrowserStack, BrowserBash reports the pass/fail status back to the grid's session, so the verdict shows correctly in the vendor dashboard.
Recording, replay, and dashboards
Every run is kept in a private on-disk store (~/.browserbash/runs, secrets masked, capped at ~200 runs). Two ways to inspect them:
# Local dashboard — free, no account, fully local (nothing leaves your machine)
browserbash dashboard # serves http://localhost:4477
browserbash run "..." --record --dashboard # run, then open the dashboard on this run
browserbash dashboard --clear # wipe the local store# Cloud dashboard — optional, opt-in per run (history across machines, shareable links)
browserbash connect --key bb_... # one-time key from browserbash.com/dashboard
browserbash run "..." --record --upload # push THIS run only
--record captures screenshots and a session video on any engine; the builtin engine additionally produces a Playwright trace you can open in the Playwright trace viewer. Without --upload, nothing is sent to the cloud. Cloud runs are retained 15 days on the free tier (extended retention is the only paid option; the CLI is fully free and open-source with no locked features).
The step fails the build automatically because the exit code carries the verdict. Use --agent so the logs are machine-readable, store secrets in the CI secret store (never in the test file), and set a --timeout so a stuck agent cannot hang the pipeline.
Best practices
Describe intent, assert outcomes. Write objectives as a user goal ("complete checkout as a standard user") and always include an explicit verification step ("Verify the order-confirmation heading is visible"). An objective with no assertion can "pass" without proving anything.
Prefer *_test.md over inline objectives for anything that recurs. Committed Markdown tests are reviewable in PRs, diff cleanly, and produce a Result.md audit trail. Reserve run "..." for exploration and one-offs.
Default to the free local stack for fast/smoke tests. Local Chrome + Ollama costs nothing; reserve paid LLMs and cloud grids for cross-browser/real-device coverage where they earn their cost.
Right-size the model. Use the smallest model that completes your flows reliably; bump to a Qwen3 / Llama-3.3-70B-class (or a hosted Claude) model only when small models stumble on multi-step objectives.
Always mark credentials as secrets. Wrap them as {"value":"...","secret":true} so they are masked in logs and NDJSON. Keep real values in CI secrets or ~/.browserbash/variables, not in committed files.
Use --agent everywhere automated. In CI and when another agent invokes BrowserBash, rely on NDJSON + the exit code for the verdict — never scrape human output.
Set --timeout and --max-steps. Natural-language agents can wander; bounding steps and time prevents runaway runs and runaway token spend.
Match engine to provider. Browserbase/local/CDP can use the default Stagehand engine; LambdaTest/BrowserStack run on the builtin engine and therefore need an Anthropic-compatible backend. Don't expect a local-Ollama-only setup to drive those grids.
Record on failure-prone or flaky flows.--record gives video (and a Playwright trace on the builtin engine) so you can see exactly what the agent did when a run is unexpectedly red.
Factor shared steps with @import. Put login and setup in a helper Markdown file and import it, so a UI change to login is fixed in one place.
Common pitfalls
Confusing BrowserBash with Browserbase / Browserless. They are different products with similar names. BrowserBash is the CLI; Browserbase is one optional provider it can target. Verify which one a command or doc refers to.
Expecting local-only models to drive cloud grids. LambdaTest and BrowserStack use the builtin (Anthropic) engine — without ANTHROPIC_API_KEY (or an ANTHROPIC_BASE_URL gateway) those runs error.
Tiny models on multi-step flows. Sub-8B models frequently misinterpret multi-step objectives, producing flaky passes/fails. Treat non-determinism here as a model-size problem first.
Hardcoding secrets in committed test files. Without {"secret":true} (or by inlining a literal), credentials leak into logs and the on-disk run store. Always parameterize and mark secrets.
Parsing human output in CI. The human reporter format can change; rely on --agent NDJSON and the exit code instead.
No timeout / no max-steps. An ambiguous objective can loop or stall; an unbounded run wastes time and tokens. Always bound automated runs.
Vague objectives without an assertion. "Go to the dashboard" with no verification can succeed while the app is actually broken. State what must be true at the end.
Assuming exact flag/model strings never change. This is an actively developed young tool. Run browserbash --help / browserbash <command> --help and check the official docs (browserbash.com/learn) when an exact spelling matters.
Forgetting --upload is opt-in (and the inverse). Nothing reaches the cloud dashboard without --upload; conversely, do upload sensitive runs only deliberately. Local history is always kept on disk regardless.
Skipping browserbash init. Running without the ./.browserbash/ workspace means no project-scoped variables, config, or test conventions — scaffold first so variable resolution and defaults work as expected.
Summary
BrowserBash lets you (or an AI coding agent) automate and test the web in plain English: write a one-shot objective or a committable *_test.md, run it on local Chrome or a cloud grid, drive it with a free local model or any cloud LLM, mask secrets, and consume NDJSON results with CI-grade exit codes. Treat it as a natural-language layer on top of Playwright: describe intent, assert outcomes, bound the run, keep credentials secret, and verify exact command/model strings against --help and the official docs since the project is evolving.
