Atlas works with Playwright through Microsoft's Playwright MCP server, which drives the browser through the accessibility tree rather than screenshots, so Atlas gets structured elements, walks a real flow, and writes the spec from what it saw. Add it with atlas mcp add playwright -- npx -y @playwright/mcp@latest --headless --isolated, then ask Atlas to open your dev server and walk the signup flow end to end.
Why Playwright MCP uses the accessibility tree, not pixels
Playwright MCP navigates the accessibility tree instead of pixels, so Atlas reasons over structured elements rather than screenshots. That 1 design choice is why Microsoft's Playwright MCP server produces reliable specs: Atlas sees a button with a role and a name, not a rectangle of colored pixels it has to guess at.
Screenshot-driven browser agents share a well known failure mode. They read a rendered image, infer what an element probably is, then emit a brittle locator that breaks the moment the layout shifts. Microsoft's Playwright MCP server avoids that by handing Atlas the accessibility tree, the browser's own structured description of the page. Roles, accessible names, and states arrive as data, not as pixels. The practical consequence is that the Playwright spec Atlas writes uses the same role-based locators a careful human would choose, because those accessible roles and names are the identifiers Atlas was reasoning over the whole time it walked the browser.
Adding Microsoft's Playwright MCP server to Atlas
Add Microsoft's Playwright MCP server with atlas mcp add playwright -- npx -y @playwright/mcp@latest --headless --isolated. Those 2 flags matter: --headless keeps the browser off your screen during an agent run, and --isolated gives every run a fresh browser profile so state from a previous run cannot leak in.
Read the command in two halves. Everything before the double dash registers a server named playwright with Atlas. Everything after it is the process Atlas launches, npx -y @playwright/mcp@latest, with its flags. The -y keeps npx from prompting, which matters because the agent starts this process rather than you. Pinning @latest means you get Microsoft's current Playwright MCP server on each launch. Start with --headless and --isolated as your defaults and only drop them when you have a specific reason, such as wanting to watch a run happen in a visible browser window.
Choosing the browser engine and the env var equivalents
The --browser flag picks the engine, taking chrome, firefox, webkit, or msedge. The same 2 knobs also exist as environment variables, PLAYWRIGHT_MCP_BROWSER and PLAYWRIGHT_MCP_HEADLESS, which is how you configure Microsoft's Playwright MCP server in CI where editing a command line is awkward.
Cross-browser coverage is the reason the flag exists. If a bug only reproduces in webkit, point the Playwright MCP server at webkit with --browser webkit and let Atlas walk the flow there, so the spec it writes is grounded in the engine where the failure actually lives. The environment variable form, PLAYWRIGHT_MCP_BROWSER and PLAYWRIGHT_MCP_HEADLESS, does the same job without changing the invocation. That is the form to prefer in a pipeline, where the same command runs everywhere and the environment is the thing that varies between jobs.
The daily workflow: walk the flow, then write the spec
The core Playwright loop with Atlas is 1 prompt. Ask Atlas to open your dev server, walk the signup flow end to end, and write the resulting spec file. Because Playwright MCP hands Atlas the accessibility tree, the spec it writes describes the elements it actually interacted with.
Writing a Playwright end to end spec by hand is mostly the tedium of discovering locators. Atlas removes that step by walking the browser first. Point Atlas at your dev server, describe the signup flow the way a user would, filling the form, submitting it, landing on the dashboard, and let Microsoft's Playwright MCP server drive the browser through it. Atlas observes the accessibility tree at each navigation step. The Playwright spec file that comes out the other side records a browser flow that genuinely worked, not a guess at one. Review that spec as a diff before it lands.
Reading failures back with --reporter=json
Let Atlas run npx playwright test --reporter=json so it reads structured failures back into context. In 2026 a JSON reporter turns a failing Playwright run into data the agent can parse precisely, instead of console text it has to scrape and can easily misread.
The red-green loop only closes if the agent can tell exactly what failed. Running npx playwright test --reporter=json gives Atlas the failing spec, the failing assertion, and the error, all as structured fields rather than as a wall of terminal output. Atlas reads that JSON, forms a hypothesis, edits the spec or the application code, and re-runs. Each cycle is grounded in a machine-readable result rather than in an approximate reading of scrolled-past text, which is what makes the iteration converge instead of wandering.
Adding vision and PDF tools with --caps
Add --caps vision,pdf when you want screenshot and PDF tools on top of the default set. Microsoft's Playwright MCP server ships a default toolset built around the accessibility tree, and those 2 extra capabilities stay opt-in for the rare task that genuinely needs to see a picture.
Keep the default toolset for most work, because the accessibility tree is the reason Atlas writes good locators, and adding screenshot tools invites the agent to reason over pixels when it does not need to. There are real exceptions. Visual regression checks and anything involving PDF output need those capabilities, and that is what --caps vision,pdf is for. Treat it as a deliberate opt-in per task rather than a permanent part of your atlas mcp add playwright command, so the default run stays structured and fast.
Setup
- 01Add Microsoft's server: atlas mcp add playwright -- npx -y @playwright/mcp@latest --headless --isolated
- 02Keep --isolated so every run gets a fresh browser profile, and use --browser chrome|firefox|webkit|msedge to pick the engine.
- 03In CI, set the same knobs as environment variables instead: PLAYWRIGHT_MCP_BROWSER and PLAYWRIGHT_MCP_HEADLESS.
- 04Ask Atlas to open your dev server, walk the signup flow end to end, and write the resulting spec file.
- 05Let Atlas run npx playwright test --reporter=json so it reads structured failures back into context.
- 06Add --caps vision,pdf only when you want screenshot and PDF tools on top of the default set.
Frequently asked questions
- how to add Playwright MCP to Atlas
- Run atlas mcp add playwright -- npx -y @playwright/mcp@latest --headless --isolated. The --isolated flag gives every run a fresh browser profile.
- does Playwright MCP use screenshots or the accessibility tree
- Playwright MCP navigates the accessibility tree instead of pixels, so Atlas reasons over structured elements rather than screenshots. That is why the locators it writes hold up.
- how do I pick the browser for Playwright MCP
- Use --browser chrome|firefox|webkit|msedge, or set PLAYWRIGHT_MCP_BROWSER in the environment. PLAYWRIGHT_MCP_HEADLESS is the env var form of --headless.
- can an AI agent write my Playwright end to end tests
- Yes. Ask Atlas to open your dev server, walk the signup flow end to end, and write the resulting spec file from the structured elements it saw.
- how does Atlas read Playwright test failures
- Let Atlas run npx playwright test --reporter=json. The JSON reporter gives it structured failures to read back into context instead of scraped console text.
- how do I enable screenshots in Playwright MCP
- Add --caps vision,pdf to the server command when you want screenshot and PDF tools on top of the default accessibility-tree toolset.
- how do I stop Playwright MCP runs from sharing browser state
- Pass --isolated. It gives every run a fresh browser profile, so cookies and storage from a previous run cannot leak into the next one.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas with IBM Granite 4.0 H Micro in 2026
IBM Granite 4.0 H Micro in Atlas, 2026: a hybrid Mamba-Transformer model at $0.017/$0.112 per Mtok on Cloudflare Workers AI, the cheapest input in the registry.
Atlas with Mixtral 8x22B: The Largest Open MoE of Its Era in 2026
Mixtral 8x22B scaled the MoE idea in April 2024: 64,000 tokens at $2.00 / 1M input tokens and $6.00 / 1M output tokens. Atlas setup, self-hosting, and honest limits.
Atlas with GLM-4.7 Flash: A Free 200K Context Model for the small_model Slot (2026)
GLM-4.7 Flash is free at $0 / $0 per Mtok with a 200,000 token context. Set it as Atlas's small_model so titles and summaries cost nothing at all.
Atlas with Mistral Medium 3.1 (2508): Setup, Cost, and Fit in 2026
Run Atlas on Mistral Medium 3.1 (2508): a 262,144 token window at $0.40 / 1M input tokens and $2.00 / 1M output tokens. Setup steps, costs, and honest tradeoffs.
Atlas with DeepSeek R1 (0528): The Open Reasoning Trace, 2026
Run Atlas on DeepSeek R1 (0528) in 2026. DeepInfra hosts the MIT-licensed open reasoning model at $0.50 per Mtok in, $2.15 per Mtok out, 160K tokens context.
Atlas for Axum in 2026
Atlas is a terminal-native AI coding agent for Axum in 2026. It decodes tower trait-bound errors, adds IntoResponse types, and runs cargo nextest run.
Atlas with North Mini Code in 2026: A 64,000 Token Output Budget
North Mini Code gives Atlas a 256,000 token context and a 64,000 token output, 8x Command A's 8,000, listed at $0 per Mtok on both input and output in the registry.
Atlas with DeepInfra: The Cheapest Open-Weights Host for an Agent Loop in 2026
Run Atlas on DeepInfra: GPT OSS 120B at $0.037/$0.17 per Mtok, DeepSeek V4 Flash at a 1,048,576 token window for $0.09 input. Setup, limits, and cost math.