> ## Documentation Index
> Fetch the complete documentation index at: https://piyushvyas.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# BAP vs Playwright

> Feature matrix and benchmark comparison between BAP and Playwright MCP/CLI for AI agent browser automation.

## Overview

BAP and Playwright both use the [Playwright](https://playwright.dev/) engine for browser automation. They differ in architecture, API design, and optimization for AI agent workflows.

| Dimension        | BAP                                                             | Playwright                                |
| ---------------- | --------------------------------------------------------------- | ----------------------------------------- |
| **Publisher**    | [browseragentprotocol](https://github.com/browseragentprotocol) | [Microsoft](https://github.com/microsoft) |
| **License**      | Apache-2.0                                                      | Apache-2.0                                |
| **Architecture** | Two-process (WebSocket bridge)                                  | Single-process                            |
| **MCP Package**  | `@browseragentprotocol/mcp`                                     | `@playwright/mcp`                         |
| **CLI Package**  | `@browseragentprotocol/cli`                                     | `@playwright/cli`                         |

## MCP Server Comparison

<CardGroup cols={2}>
  <Card title="BAP MCP" icon="layer-group">
    * 23 MCP tools - Composite `act` batches N steps in 1 call - `observe` returns structured
      elements with refs - `extract` with JSON Schema - 6 fusion operations - Response tiers
      (full/interactive/minimal) - WebMCP discovery - Session persistence across restarts
  </Card>

  <Card title="Playwright MCP" icon="play">
    * 31 tools (17 core + 6 vision + 5 test + 3 other) - No built-in batching - `browser_snapshot`
      returns raw accessibility tree - `browser_evaluate` for custom JS extraction - Lower per-call
      latency (single-process) - Simpler setup (`npx @playwright/mcp`) - 27.5k GitHub stars,
      Microsoft-backed
  </Card>
</CardGroup>

## Feature Matrix

| Feature                           |                  BAP                  |                    Playwright MCP                   |
| --------------------------------- | :-----------------------------------: | :-------------------------------------------------: |
| Composite actions (batch N steps) |                  Yes                  |                          No                         |
| Structured observation with refs  |                  Yes                  |                  Accessibility tree                 |
| JSON Schema extraction            |                  Yes                  |                      Custom JS                      |
| Navigate + observe fusion         |                  Yes                  |                          No                         |
| Act + pre/post observe fusion     |                  Yes                  |                          No                         |
| Incremental observation           |                  Yes                  |                          No                         |
| Response compression tiers        |                  Yes                  |                          No                         |
| WebMCP tool discovery             |                  Yes                  |                          No                         |
| Session persistence               |                  Yes                  |                          No                         |
| Multi-client (CLI + MCP)          |                  Yes                  |                          No                         |
| Batched form filling              |    `act` (N fills + click = 1 call)   | `browser_fill_form` (fills batched, click separate) |
| Per-call latency                  |         +50-200ms (WebSocket)         |                  Lower (in-process)                 |
| Element disambiguation            | Semantic selectors (can be ambiguous) |           Positional refs (always unique)           |
| SKILL.md support                  |              13 platforms             |             Claude Code, GitHub Copilot             |

## CLI Comparison

| Dimension              | BAP CLI                                 | Playwright CLI                    |
| ---------------------- | --------------------------------------- | --------------------------------- |
| **Commands**           | 26                                      | \~70+ (granular)                  |
| **Composite actions**  | `bap act fill:...=val click:...`        | Individual commands               |
| **Semantic selectors** | `role:button:"Submit"`, `label:"Email"` | Accessibility tree refs           |
| **Observation**        | `bap observe --tier=interactive`        | `playwright-cli snapshot`         |
| **Extraction**         | `bap extract --fields="title,price"`    | `playwright-cli eval` (custom JS) |

## Benchmark Results

All data from the [reproducible benchmark suite](https://github.com/browseragentprotocol/benchmarks). See [Benchmarks](/comparisons/benchmarks) for full methodology.

| Scenario      | BAP Standard | BAP Fused | Playwright | Std vs PW | Fused vs PW |
| ------------- | :----------: | :-------: | :--------: | :-------: | :---------: |
| baseline      |       2      |     2     |      2     |    Tie    |     Tie     |
| observe       |       2      |     1     |      2     |    Tie    |     -50%    |
| extract       |       2      |     2     |      2     |    Tie    |     Tie     |
| form          |       4      |     3     |      5     |    -20%   |     -40%    |
| **ecommerce** |     **8**    |   **5**   |   **11**   |  **-27%** |   **-55%**  |
| workflow      |       5      |     4     |      5     |    Tie    |     -20%    |
| **Total**     |    **23**    |   **17**  |   **27**   | **\~15%** |  **\~37%**  |

<Note>
  The fair comparison is **BAP Standard vs Playwright**. BAP Standard follows the same
  observe-then-act pattern as Playwright. BAP Fused is an optimization layer that is not
  apples-to-apples.
</Note>

## Where BAP Wins

* **Composite `act`**: Batching multiple steps into one call is the primary advantage. Most impactful in multi-step flows like ecommerce (8 vs 11 calls).
* **Fused operations**: `navigate(observe:true)` and `act(postObserve:true)` eliminate redundant roundtrips.
* **Structured `extract`**: JSON Schema-based extraction vs writing custom JS.
* **Session persistence**: Browser state survives across client restarts.
* **Multi-client**: CLI and MCP share the same browser simultaneously.

## Where Playwright Wins

* **Per-call latency**: Single-process architecture. No WebSocket overhead.
* **Element disambiguation**: Positional snapshot refs uniquely identify elements. BAP's observe can return ambiguous selectors for identical elements (e.g., 6 "Add to cart" buttons).
* **Setup simplicity**: `npx @playwright/mcp` -- single process, no daemon management.
* **Ecosystem**: 27.5k GitHub stars, Microsoft-backed, extensive testing integration.
* **Tool count**: 31 tools covering more granular operations.

## When to Use Which

<CardGroup cols={2}>
  <Card title="Choose BAP" icon="check">
    * Multi-step form automation - AI agent workflows (observe-decide-act loop) - Need CLI + MCP
      access to same browser - Session persistence matters - Token efficiency is critical
  </Card>

  <Card title="Choose Playwright" icon="check">
    * Per-call latency is the priority - Already in the Playwright testing ecosystem - Need the
      simplest possible setup - Element disambiguation is critical - Vision-based automation (6 vision
      tools)
  </Card>
</CardGroup>