Turn any AI browser run into clean, production Playwright code.

recast is a compiler that takes AI browser agent recordings and emits clean, readable, static Playwright test code - with no LLM required at replay time, no proprietary runtime dependencies, and no magic.

workflow-use JSON  ──┐
HAR file           ──┤  recast compile  ──▶  playwright.spec.ts
CDP event log      ──┤                  ──▶  .env.example
MCP tool call log  ──┘

Why

AI browser agents (workflow-use, browser-use, Skyvern, Operator) re-reason through browser tasks at runtime — burning LLM tokens on every step, every run, forever. Recorded workflows are locked to proprietary runtimes and can't be committed to CI pipelines, reviewed by developers, or diff-ed.

recast fills the gap: compile once, run forever, on plain Playwright.

Demo

$ recast compile login_workflow.json

recast: INFO  detected format: workflow-use JSON
recast: INFO  parsed 5 steps
recast: INFO  [dedup] removed 0 duplicate steps
recast: INFO  [selector] hardened 2 selector(s)
recast: INFO  [waits] injected 3 explicit wait(s)
recast: INFO  [sanitize] 2 credential(s) replaced with environment variables
recast: INFO  [branch] no conditional branches detected
recast: INFO  [emit] wrote playwright-ts to ./recast-out/test_login_workflow.spec.ts
recast: INFO  [emit] wrote .env.example to ./recast-out/.env.example

$ cat ./recast-out/test_login_workflow.spec.ts

import { test, expect } from '@playwright/test';

test('login_workflow', async ({ page }) => {
  await page.goto('https://app.example.com/login');
  await page.waitForLoadState('networkidle');

  await page.locator('#email').fill(process.env.TEST_EMAIL!);
  await page.locator('#password').fill(process.env.TEST_PASSWORD!);

  await page.getByRole('button', { name: 'Sign in' }).click();
  await page.waitForNavigation({ timeout: 10000 });

  await page.locator('.dashboard').waitFor({ state: 'visible', timeout: 30000 });
});

Install

Download a pre-built binary (no Go required):

# macOS arm64 (Apple Silicon)
curl -L https://github.com/yagna-1/recast/releases/latest/download/recast_darwin_arm64.tar.gz | tar xz
sudo mv recast /usr/local/bin/

# macOS x86_64 (Intel)
curl -L https://github.com/yagna-1/recast/releases/latest/download/recast_darwin_x86_64.tar.gz | tar xz
sudo mv recast /usr/local/bin/

# Linux x86_64
curl -L https://github.com/yagna-1/recast/releases/latest/download/recast_linux_x86_64.tar.gz | tar xz
sudo mv recast /usr/local/bin/

Or grab the .zip for Windows from the releases page.

Go install:

go install github.com/yagna-1/recast/cmd/recast@latest

Build from source:

git clone https://github.com/yagna-1/recast
cd recast
go mod tidy
make build
./bin/recast version

Quickstart

# Compile a workflow-use JSON recording to Playwright TypeScript
recast compile workflow.json

# Output:
# recast: INFO  detected format: workflow-use JSON
# recast: INFO  parsed 6 steps
# recast: INFO  [sanitize] 2 credential(s) replaced with environment variables
# recast: INFO  [emit] wrote playwright-ts to ./recast-out/test_login_and_download.spec.ts
# recast: INFO  [emit] wrote .env.example to ./recast-out/.env.example
# Exit code can be 1 when warnings are present (partial success).

# View output
cat ./recast-out/test_login_and_download.spec.ts

Commands

recast compile

Compile a workflow to Playwright test code.

recast compile [flags] <input-file>

Flags:
  -o, --output string       Output directory (default "./recast-out/")
  -t, --target string       playwright-ts | playwright-py | ir-json (default "playwright-ts")
      --no-optimize         Skip optimizer passes
      --no-harden           Skip selector hardening
      --inject-assertions   Add post-action assertions
      --replay-exact        Preserve recorded sequence; disable behavior-changing heuristic passes
      --strict              Exit 2 on any warning
  -v, --verbose             Detailed output

Examples:

recast compile workflow.json
recast compile workflow.json -t playwright-py -o ./tests/
recast compile trace.har --inject-assertions
recast compile mcp-log.jsonl -t ir-json

recast validate

Validate an input file without compiling.

recast validate workflow.json

recast verify

Verify compiled Playwright output against the original trace using screenshot-free signals (URL + key selectors).

recast verify ./recast-out/test_login_and_download.spec.ts --against workflow.json
# optional runtime check (no screenshot diff)
recast verify ./recast-out/test_login_and_download.spec.ts --against workflow.json --runtime
# strict runtime mode (fails if runtime cannot execute or test fails)
recast verify ./recast-out/test_login_and_download.spec.ts --against workflow.json --runtime --runtime-strict --runtime-timeout-sec 120

Notes:

• If the provided test file path is slightly off, recast verify auto-resolves to the closest *.spec.ts/*.spec.py file in the same directory and prints a warning.
• Runtime verification injects safe placeholder defaults for sanitized variables (TEST_EMAIL, TEST_PASSWORD, RECAST_VAR_1..10) when unset, so strict checks can run in clean environments.

recast formats

List all supported input formats and output targets.

recast version

Print version information.

Exit codes

• 0 success
• 1 partial success with warnings
• 2 hard validation/compilation failure
• 3 input file not found/unreadable
• 4 unsupported input/output format

What recast does

1. Ingestion

Parses input from supported formats into a normalized Intermediate Representation (IR). The IR is the AST — every optimization and code generation step operates on it.

Supported input formats:

• workflow-use JSON ({"workflow_name": ..., "steps": [...]})
• HAR (HTTP Archive) — exported from any browser DevTools
• CDP Event Log — raw Chrome DevTools Protocol logs
• MCP Tool Call Log — from agents using MCP browser tools

2. Optimizer passes

Pass	Name	Default
1	Deduplication — remove consecutive identical actions	✓
2	Selector stabilization — upgrade fragile CSS to ARIA/role	✓
3	Wait inference — inject explicit waits	✓
4	Credential sanitization — replace secrets with env vars	✓ always
5	Branch detection — flag missing conditional branches	✓
6	Assertion injection — add post-action assertions	opt-in

3. Emit

Generates clean, readable test code in the target language.

Supported targets:

• playwright-ts — Playwright TypeScript (.spec.ts)
• playwright-py — Playwright Python (.py)
• ir-json — Normalized IR JSON (for debugging / community tools)

Input → Output Example

Input (workflow.json):

{
  "workflow_name": "login_and_download",
  "steps": [
    { "type": "navigate", "url": "https://app.example.com/login" },
    { "type": "fill", "selector": "#email", "value": "user@example.com" },
    { "type": "fill", "selector": "#password", "value": "hunter2" },
    { "type": "click", "selector": "button[type=submit]" },
    { "type": "wait_for", "selector": ".dashboard" }
  ]
}

Output (test_login_and_download.spec.ts):

import { test, expect } from '@playwright/test';

test('login_and_download', async ({ page }) => {
  // Navigate to login page
  await page.goto('https://app.example.com/login');
  await page.waitForLoadState('networkidle');

  // the email input field
  await page.locator('#email').fill(process.env.TEST_EMAIL!);

  // the password input field
  await page.locator('#password').fill(process.env.TEST_PASSWORD!);

  // the Sign in button
  await page.locator('button[type=submit]').click();
  await page.waitForNavigation({ timeout: 10000 });

  await page.locator('.dashboard').waitFor({ state: 'visible', timeout: 30000 });
});

.env.example:

# Generated by recast — fill in actual values before running tests
# DO NOT commit this file with real values

# step_002: fill value was sanitized (detected: email pattern)
TEST_EMAIL=

# step_003: fill value was sanitized (detected: high-entropy string)
TEST_PASSWORD=

Security

Credential sanitization runs always and cannot be disabled. Any fill step value that matches an email pattern, common credential field name, JWT token, credit card number, or high-entropy string is automatically replaced with an environment variable reference.

The .env.example file lists all sanitized variables. Fill in the real values locally; never commit them.

Architecture

recast is structured identically to a programming language compiler:

Input Formats      →  Ingestion Adapters  →  IR (AST)  →  Optimizer Passes  →  Emitters  →  Output

Adding a new input format = write one ingestion adapter. Adding a new output target = write one emitter. Optimizer passes run on IR regardless of input or output.

The IR package (recast-ir) is published as a standalone Go module so agent frameworks can emit IR directly — bypassing file-based ingestion entirely.

Development

git clone https://github.com/yagna-1/recast
cd recast
go mod tidy     # fetch cobra, viper, testify (populates go.sum)
make test       # all unit tests
make build      # compile binary
make golden-update  # generate golden files on first run
make golden-test    # run golden output tests (requires golden-update first)

CI

GitHub Actions runs:

• go mod tidy consistency check
• gofmt format check
• go vet
• make test
• make golden-test
• make build

See .github/workflows/ci.yml.

Contributing

See CONTRIBUTING.md for development workflow and PR expectations.

First-time contributor? Look for issues labeled good first issue — each one has a step-by-step guide.

Project structure

recast/
├── cmd/recast/          # CLI entrypoint
├── recast-ir/           # IR type definitions, builder, validator, marshaler
├── internal/
│   ├── ingestion/       # Format adapters (workflow-use, HAR, CDP, MCP)
│   ├── optimizer/       # Passes 1-6
│   ├── emitter/         # Playwright TS, Python emitters
│   └── config/          # Config management
├── pkg/recast/          # Public Go library API
└── testdata/            # Fixtures and golden files

Code of Conduct

This project follows the Contributor Covenant Code of Conduct. By participating, you agree to uphold these standards.

License

MIT — see LICENSE.