BetterWright keeps one browser session open across an agent's calls, so the agent can navigate, interact with pages, and read results without starting a new browser each time. The agent drives it with Playwright JavaScript snippets that run in a sandboxed worker and come back as result envelopes reporting success or an error.
Instead of dumping whole pages into the context window, BetterWright returns compact accessibility-tree snapshots with element references, interactive-only views, scoped subtrees, and diffs that show only what changed since the last snapshot. Typed input is read back and the step fails if the text did not land. Proof screenshots are saved as artifacts that the result cites rather than inlined pixels.
Existing agents connect through a CLI with packaged skills, a stdio Model Context Protocol (MCP) server, or a JavaScript and TypeScript SDK. An optional built-in agent accepts a natural-language task and runs the browser steps itself, using a model you configure: a ChatGPT subscription through Codex sign-in, Cerebras, compatible endpoints, or a local model that setup selects for the available GPU. The browser runtime itself needs no model.
The default browser is a managed Chromium build called BetterChromium. You can also supply a local Chromium executable, attach over the Chrome DevTools Protocol, use one of several cloud browser providers, or embed it through an Electron adapter. Network access is permissive by default, with flags to block private and loopback destinations and a fixed block on cloud metadata endpoints for locally launched browsers.
Features
- Persistent sessions: named sessions keep tabs and page state between calls, and profiles keep cookies for separate accounts
- Compact snapshots: accessibility-tree views with element references, interactive-only filtering, scoped subtrees, and change-only diffs
- Verified input: typing reads the field back and fails the step when the text is missing
- Proof screenshots: screenshots saved as artifacts that the result references
- Credential vault: encrypted stored credentials, filled only on matching URLs and redacted from results
- Human handoff: live view of the browser and handoff for steps such as multi-factor authentication
- Network policy: optional blocking of private and loopback destinations, plus a fixed cloud metadata block
- Agent interfaces: CLI with installable skills, an MCP server, and a JavaScript and TypeScript SDK
- Built-in agent: natural-language task execution with a configured hosted or local model
- Browser providers: managed BetterChromium, a local Chromium, CDP attachment, cloud browsers, or an Electron host
- Ad blocking: Ghostery ad and tracker blocking on by default
