Playwright Report Mcp

Playwright Report MCP

An MCP (Model Context Protocol) server for running Playwright tests and reading structured results, failed test details, and attachment content — designed for AI agents doing test failure analysis.

[License: MIT] [Node.js: 22+]

Table of contents

• What it is
• What it is NOT
• Why
• Quick start
• Tools
• Attachments
• Installation
• Configuration
• Requirements
• Troubleshooting
• Development
• Cutting a release
• Contributing
• Release
• License

What it is

Playwright Report MCP gives an AI agent structured, token-efficient access to Playwright test outcomes. It runs your test suite, reads the JSON reporter output, and surfaces exactly what the agent needs: which tests failed, what the errors were, and the content of relevant attachments.

What it is NOT

There are many Playwright MCP servers that control a browser — they navigate pages, click elements, fill forms, and take screenshots. Playwright Report MCP is not one of those.

Why

The problem with existing approaches

Default reporters (list / dot) — Playwright's default reporters print human-readable output to stdout. Compact, but lossy: no attachment paths, no retry breakdown, no structured data.

HTML reporter (report.html) — A self-contained SPA bundle (typically 2–50 MB). Not machine-readable as text and exceeds any LLM context window.

Reading results.json directly — Works, but a full JSON report for even a small test suite is 10,000–20,000 tokens. For a failing test, most of that is passing test metadata you don't need.

What Playwright Report MCP does instead

• Filters results.json to only failed tests
• Returns structured, typed JSON the agent can act on immediately
• Exposes individual attachments by name so the agent fetches only what it needs
• Works on results produced by anyone — CI pipeline, a human, or the agent itself

Token cost comparison (one failed test in a 20-test suite)

Approximate input token counts based on Claude tokenization (~3–4 characters per token for mixed JSON/text content).

¹ Custom text attachments — e.g. AI diagnosis (~500–2,000 tokens) and console logs (~200–500 tokens) added via testInfo.attach() in your own fixtures.

² Full page HTML snapshot — a custom fixture that attaches the full rendered page HTML on failure. Large pages alone can reach 30,000–80,000 tokens and dominate cost regardless of whether MCP is used.

Key observations:

• For a live run, stdout (list/dot) is compact but gives the agent no path to attachment content — dead end for deeper analysis
• Reading results.json directly costs ~12,500–23,000 tokens even when only one test failed — most of it is passing test metadata the agent doesn't need
• The biggest MCP gains are in the middle rows: getting error messages + page state from existing results at ~4–45× lower token cost
• Full page HTML snapshot dominates cost either way; skipping it in favour of error-context is the single largest optimisation available

CI failure analysis

The primary use case: your CI pipeline runs the tests, the agent picks up the results after the fact and diagnoses failures. get_failed_tests reads results.json regardless of who triggered the run. No re-run needed.

Quick start

1. Install via npx (recommended)

No clone or build step needed — npx downloads and runs the server automatically:

{  "mcpServers": {    "playwright-report-mcp": {      "command": "npx",      "args": ["-y", "playwright-report-mcp"],      "type": "stdio"    }  }}

Or build from source:

git clone https://github.com/hubertgajewski/playwright-report-mcp.gitcd playwright-report-mcpnpm install && npm run build

2. Add the JSON reporter to your Playwright project

// playwright.config.tsreporter: [  ['json', { outputFile: 'test-results/results.json' }],  ['html'], // keep any existing reporters],

3. Register in .mcp.json

{  "mcpServers": {    "playwright-report-mcp": {      "command": "npx",      "args": ["-y", "playwright-report-mcp"],      "type": "stdio"    }  }}

4. Ask your AI agent

Run the Playwright tests and tell me what failed.

Compatibility

Tested with Claude Code (CLI). Should work with any MCP-compatible client that supports stdio transport, including Claude Desktop, Cursor, Cline, Windsurf, and Continue.dev — but these have not been verified.

The stdio server supports both MCP protocol eras from one entrypoint:

• Modern: protocol revision 2026-07-28, selected by clients using version negotiation (for example, versionNegotiation: { mode: "auto" }).
• Legacy: supported 2025-era revisions, selected by clients that use the traditional initialize handshake. This remains the default behavior in the MCP client SDK.

The opening exchange pins one era for the connection lifetime. A client that pins an unsupported revision receives an explicit negotiation error; the server does not silently switch it to another era.

Tools

The project-scoped tools accept an optional workingDirectory parameter — see Multi-worktree support. get_run_status can use either a runId from run_tests with wait: false, or a workingDirectory lookup for the latest tracked run.

run_tests

Runs the Playwright test suite and returns structured pass/fail results.

Returns: exit code, run stats, and a summary of all tests with status, duration, and error per project.

When wait is false, returns immediately with runId, process metadata, compact numeric progress, and current results.json status. Poll get_run_status with that runId until state is completed, failed, or timedOut. The server parses Playwright progress markers such as [528/662] from stdout and discards raw stdout/stderr text so repeated polling stays token-efficient. The server allows one active tracked run per working directory, caps active tracked runs globally, and keeps a bounded history of recent terminal runs, so very old runId values can expire.

get_run_status

Returns the current status for a non-blocking run started by run_tests with wait: false.

If both fields are omitted, workingDirectory defaults to ".". If no run is tracked for the resolved directory, the tool returns state: "idle" plus results.json metadata and last parsed stats when readable. It does not process-scan for external npx playwright test commands that were not started through this MCP server.

Returns: run state, tracking flag, pid, timestamps, elapsed duration, timeout, command metadata (executable, args, cwd), optional envKeys (override names only — never values), progress: { current, total }, exit code, signal, spawn/timeout error when present, results.json path/existence/mtime/size/freshness, and parsed report stats when the report was updated after the run started. When progress has not appeared yet, current and total are null; when a terminal run has readable final stats, progress is set to the derived completed total.

get_failed_tests

Returns failed tests from the last run with error messages and attachment paths. Does not re-run tests — reads the existing results.json.

Returns: failed test count, titles, file paths, per-project status, error messages, and attachment paths.

get_test_attachment

Reads the content of a named text attachment for a specific test from the last run.

Returns: the attachment content as text. Binary attachments and files over 1 MB are rejected with an error. Attachment paths recorded in results.json that escape workingDirectory (via .. or absolute paths pointing elsewhere) are refused.

list_tests

Lists all tests with their spec file and tags without running them.

Attachments

Playwright attaches files to failed tests automatically. get_test_attachment can read any text attachment by name.

The error-context attachment is the most useful for projects without custom fixtures — it gives a semantic, structured view of the page at the moment of failure with no setup required.

Installation

Via npx (recommended) — use the npx config shown in Quick start. No local installation needed.

From source:

git clone https://github.com/hubertgajewski/playwright-report-mcp.gitcd playwright-report-mcpnpm installnpm run build

Configuration

Add to your .mcp.json at the root of your project:

{  "mcpServers": {    "playwright-report-mcp": {      "command": "npx",      "args": ["-y", "playwright-report-mcp"],      "type": "stdio"    }  }}

Environment variables

Set PW_RESULTS_FILE if your playwright.config.ts writes the report to a non-default location. Leave it unset in multi-worktree setups so each workingDirectory gets its own test-results/results.json.

Multi-worktree support

run_tests, list_tests, get_failed_tests, and get_test_attachment all accept an optional workingDirectory parameter — absolute, or relative to the MCP server's launch directory. get_run_status also accepts workingDirectory when runId is omitted; when runId is supplied, any supplied workingDirectory must resolve to that run's recorded directory. This lets a single long-lived MCP session drive tests across multiple git worktrees without restarting.

Because a Playwright config is a Node module that executes on playwright test startup, the server guards the parameter with an allowlist. Callers that point workingDirectory at a directory outside PW_ALLOWED_DIRS get a structured error and no child process is spawned.

Default (no worktrees). Leave PW_ALLOWED_DIRS unset. The allowlist becomes "." — only the launch directory — and the default workingDirectory (also ".") resolves to the launch directory. Zero configuration.

Sibling worktrees. Set PW_ALLOWED_DIRS=".." in your .mcp.json to authorize every sibling of the launch directory. Relative entries resolve against the launch cwd at startup, so the same .mcp.json works for every contributor without baking in absolute paths:

{  "mcpServers": {    "playwright-report-mcp": {      "command": "npx",      "args": ["-y", "playwright-report-mcp"],      "env": { "PW_ALLOWED_DIRS": ".." },      "type": "stdio"    }  }}

Then point calls at any sibling worktree:

{  "name": "run_tests",  "arguments": { "workingDirectory": "../my-app-feat-auth" },}

Multiple projects. Either launch the MCP client from each project and use the default allowlist, or set PW_ALLOWED_DIRS to the shared parent and pass workingDirectory per call. The allowlist check runs at a path-segment boundary, so an entry authorizing /src/my-app will not authorize /src/my-app-evil.

Breaking change (2.x → next): the PW_DIR env var has been removed. Either launch the MCP client from inside the Playwright project directory (zero-config, default workingDirectory: "." works), or pass workingDirectory per call and set PW_ALLOWED_DIRS accordingly.

Environment overrides

run_tests and list_tests accept an optional env map that is merged onto the MCP process environment before spawning Playwright. The child therefore sees operator-allowlisted overrides on top of launch env and the parent process. Playwright configs that load dotenv without overriding existing variables keep the tool-supplied value.

The env argument is off until the operator sets PW_ALLOWED_ENV in .mcp.json. That is the same containment model as PW_ALLOWED_DIRS: the human expands what the model may do; the model cannot widen it.

A hard denylist always wins over the allowlist (comparison is case-insensitive):

• prefixes NODE_, LD_, DYLD_, NPM_CONFIG_ (the child is launched with npx, which reads npm config from the environment)
• PATH, HOME, DOTENV_CONFIG_PATH
• proxy variables HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, NO_PROXY, FTP_PROXY
• any name matching PASSWORD, SECRET, TOKEN, API_KEY, ACCESS_KEY, CREDENTIAL, PRIVATE_KEY, AUTH_SOCK, or AUTH_CONFIG

Keys must be POSIX names (^[A-Z_][A-Z0-9_]*$). Values cannot contain a NUL and are capped at 256 characters, with at most 16 entries per call. The value of ENV is further restricted to ^[A-Za-z0-9._-]+$ so it cannot be used as a POSIX shell startup-file path (ENV=/tmp/evil.sh is rejected).

Secrets (basic auth, passwords, API keys) stay in the project .env or in MCP launch env. They are never tool arguments. get_run_status may list override names as envKeys and never echoes override values. There is no envFile tool argument.

Target environment name (ENV=staging):

{  "mcpServers": {    "playwright-report-mcp": {      "command": "npx",      "args": ["-y", "playwright-report-mcp"],      "env": {        "PW_ALLOWED_DIRS": "..",        "PW_ALLOWED_ENV": "ENV"      },      "type": "stdio"    }  }}

{  "workingDirectory": "playwright/typescript",  "env": { "ENV": "staging" }}

Other project switches (TEST_ENV or BASE_URL):

{  "mcpServers": {    "playwright-report-mcp": {      "command": "npx",      "args": ["-y", "playwright-report-mcp"],      "env": { "PW_ALLOWED_ENV": "TEST_ENV,BASE_URL" },      "type": "stdio"    }  }}

{ "env": { "TEST_ENV": "qa" } }

{ "env": { "BASE_URL": "https://stage.example.com" } }

Requirements

• Node.js 22+
• @playwright/test 1.40 or later
• JSON reporter configured in your Playwright project

Playwright's default reporters (list locally, dot on CI) write to stdout only — they produce no file that can be read after the run. Add the JSON reporter alongside whatever reporters you already use:

// playwright.config.tsreporter: [  ['json', { outputFile: 'test-results/results.json' }],  ['html'],  // keep any existing reporters  ['list'],],

Troubleshooting

No results.json found — run tests first

The JSON reporter is not configured or is writing to a different path. Verify your playwright.config.ts has ['json', { outputFile: 'test-results/results.json' }].

list_tests parsed 0 tests from non-empty output

The --list output format may have changed in your version of Playwright. Open an issue with your Playwright version and the raw stdout output.

Attachment "..." is binary and cannot be returned as text

screenshot and video attachments are binary files. Use get_failed_tests to get attachment paths and open them directly if needed.

Attachment "..." is too large to return inline

The attachment exceeds 1 MB. Read the file directly from the path returned by get_failed_tests.

Development

npm test          # run tests oncenpm run test:watch  # watch mode

Runtime code lives under src/ and compiles to dist/. Tests use Vitest with focused unit coverage for helpers plus MCP tool integration coverage via InMemoryTransport. No build step or Playwright installation required to run the regular test suite.

Cutting a release

Releases are produced by pushing a v* tag. .github/workflows/release.yml picks up the tag, verifies the tag matches all three version fields, runs npm ci + npm run build + npm test, creates a GitHub Release with auto-generated notes categorized per .github/release.yml (Features / Bug fixes / Documentation / Dependencies / Other changes), and publishes to npm. .github/workflows/publish-mcp.yml then chains off Release via workflow_run and publishes server.json to the MCP registry. Merging to main does not trigger a publish.

Version lives in three places and all three must match the tag before pushing it:

• package.json → version
• server.json → top-level version
• server.json → packages[0].version

Bump all three in one PR and merge to main before cutting the release. release.yml fails the run if the tag disagrees with any of these values.

Ritual:

# After the version-bump PR has merged to main:git checkout main && git pullgit tag v1.0.5git push origin v1.0.5# → release.yml fires: verifies tag, builds, tests, creates GitHub Release, publishes to npm# → publish-mcp.yml chains off Release and publishes server.json to the MCP registry

Flow:

1. Open a bump PR that updates all three version fields. Merge it to main.
2. Tag the bump commit v<version> and push the tag.
3. release.yml verifies tag/version alignment, runs npm ci, confirms the version is not already published on npm, runs npm run build + npm test, creates the GitHub Release, then publishes to npm with npm publish --access public --provenance.
4. publish-mcp.yml (triggered by workflow_run on Release) re-verifies the version fields, confirms the version is not already on the MCP registry, and publishes server.json to registry.modelcontextprotocol.io.

No repository secrets required. Both npm and the MCP registry authenticate via GitHub OIDC (npm trusted publishers). The trusted publisher for npm is configured on npmjs.com under the package's Settings → Publishing access → Trusted Publisher section — no NPM_TOKEN secret exists or is needed.

Recovery from a failed publish: npm refuses to republish an existing version and restricts unpublishing after 72 hours. If a publish fails for any reason, bump to the next patch version in a new PR and cut a new release — do not try to re-run the failed release.

Contributing

See CONTRIBUTING.md for bug reports, pull requests, development setup, and commit conventions.

License

MIT — Copyright (c) Hubert Gajewski