🎭 playwright-trace-decoder-mcp

An MCP server that unpacks and structures Playwright trace.zip archives so AI agents can perform root-cause analysis on CI failures — without drowning in raw JSON or blowing up the context window.

🤔 The Problem

When a Playwright test fails in CI, you get a trace.zip. It's a binary blob. LLMs can't read it natively, and dumping the raw contents exceeds the context window. Engineers end up copying log snippets into ChatGPT manually like it's 2022.

This MCP server solves that: 16 focused tools that expose exactly the signal an agent needs to diagnose a failure, with pagination and ARIA compression to keep token costs low.

🐸 E2E Failure Investigation Example

Here is a quick look at how an AI agent uses the new tools in v0.3.0 to instantly find and inspect a failure:

1. Locate the exact source code bug via map_locator_to_source:

   Copy

   // Request arguments
   { "trace_path": "/path/to/trace.zip" }
   
   // Response payload
   {
     "action_type": "Click locator('#super-toad-not-found')",
     "locator": "#super-toad-not-found",
     "error": "TimeoutError: locator.click: Timeout 5000ms exceeded.",
     "step_title": "Click locator('#super-toad-not-found')",
     "stack": [
       {
         "file": "/Users/albertdev/Projects/ideas/sample-playwright-project/tests/google-pom.spec.ts",
         "line": 18,
         "column": 17
       }
     ],
     "source_location": {
       "file": "/Users/albertdev/Projects/ideas/sample-playwright-project/tests/google-pom.spec.ts",
       "line": 18,
       "column": 17
     }
   }

   No more guessing! The agent knows exactly which file, line, and column caused the timeout.

2. Extract critical visual frames around the failure via extract_critical_frames:

   Copy

   // Request arguments
   { "trace_path": "/path/to/trace.zip", "limit": 1 }
   
   // Response payload
   [
     {
       "timestamp": 1779137404287,
       "mime_type": "image/jpeg",
       "step_title": "Clicking #super-toad-not-found element",
       "data": "/9j/4AAQSkZJRgABAQAAAQABAAD/..." // Base64 JPEG
     }
   ]

   Allows the agent to visual-verify page state immediately before/after failure without pulling massive image lists.

3. Trim the trace to save CI storage / transfer costs via trim_trace_archive:

   Copy

   // Request arguments
   { "trace_path": "/path/to/trace.zip" }
   
   // Response payload
   {
     "original_size_bytes": 2449682,
     "trimmed_size_bytes": 511698,
     "compression_ratio_percent": 79,
     "trimmed_trace_path": "/path/to/trace.trimmed.zip"
   }

   Shrinks large traces by deleting screenshots outside the critical failure window. Saved 79% of disk space!

🛠️ Tools

Tools are grouped by how an agent should sequence them when diagnosing a failure.

Inspection — read trace data

All list-returning tools support limit (1–500, default 50) and offset pagination with a has_more flag.

trace_path accepts either an absolute local path or an HTTPS URL — the server downloads the file automatically and caches it for the session.

DOM / UI analysis

Root-cause analysis

Performance analysis

💬 Suggested agent workflow

get_trace_summary              ← what failed?
get_causal_chain_for_failure   ← what led up to it?
get_aria_accessibility_tree    ← what did the page look like?
get_screenshot_at_failure      ← ARIA empty? get the actual screenshot
get_dom_mutation_delta         ← what changed right before the failure?
analyze_race_conditions        ← was a network request still pending?
correlate_dom_and_network      ← which fetch caused which DOM change?
compare_traces                 ← flaky? compare to a passing run
detect_performance_anomalies   ← timeout but no JS error? check for Long Tasks

🏗️ Architecture

trace.zip
  ├── *.trace          → JSONL: before/after action pairs, console events, frame snapshots
  ├── *.network        → JSONL: HAR resource-snapshot entries
  └── resources/
      ├── page@*.jpeg  → screenshots taken during the run
      └── ...          → fonts, stylesheets, other captured resources

The parser streams each file line-by-line (no full-buffer split) and caches results in-process with an LRU (max 50 entries), keyed by path + mtime. Re-reading the same unmodified trace costs zero I/O.

Frame snapshots store the DOM as nested arrays (["TAG", {attrs}, ...children]). The ARIA translator walks this tree and outputs compact YAML, reducing token cost by ~90% vs raw HTML.

🏗️ Stack

• @modelcontextprotocol/sdk — MCP server runtime
• adm-zip — zip extraction
• zod v4 — input schema validation
• TypeScript, ESLint, Prettier, Husky, GitHub Actions CI

📋 Scripts

npm run build        # compile TypeScript → dist/
npm run lint         # ESLint
npm run format       # Prettier --write
npm run format:check # Prettier check (used in CI)

📄 License

MIT

🚀 Setup

Build from source

git clone https://github.com/vola-trebla/playwright-trace-decoder-mcp.git
cd playwright-trace-decoder-mcp
npm install
npm run build

Add to your MCP client

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json)

{
  "mcpServers": {
    "playwright-trace-decoder": {
      "command": "node",
      "args": ["/absolute/path/to/playwright-trace-decoder-mcp/dist/index.js"]
    }
  }
}

Cursor (.cursor/mcp.json) or VS Code (.vscode/mcp.json)

{
  "mcpServers": {
    "playwright-trace-decoder": {
      "command": "node",
      "args": ["/absolute/path/to/playwright-trace-decoder-mcp/dist/index.js"]
    }
  }
}

Claude Code

claude mcp add playwright-trace-decoder \
  node /absolute/path/to/playwright-trace-decoder-mcp/dist/index.js

Docker

docker build -t playwright-trace-decoder-mcp .

{
  "mcpServers": {
    "playwright-trace-decoder": {
      "command": "docker",
      "args": ["run", "--rm", "-i", "-v", "/path/to/traces:/traces", "playwright-trace-decoder-mcp"]
    }
  }
}

💬 Example usage

Basic failure analysis

Ask your agent:

"The CI run failed. Here's the trace: /tmp/trace.zip. What went wrong and why?"

The agent calls get_trace_summary → get_causal_chain_for_failure → get_aria_accessibility_tree, drilling deeper as needed — without you copy-pasting anything.

When the page was blank or redirected

"The ARIA tree is empty. Can you show me what was actually on screen when it failed?"

The agent calls get_screenshot_at_failure and gets the JPEG taken closest to the moment of failure — useful for catching captchas, error pages, or unexpected redirects.

Flakiness diagnosis

"This test passes locally but fails in CI. Compare these two traces and tell me what was different."

The agent calls compare_traces, which LCS-aligns both action sequences and surfaces the first structural divergence, timing anomalies, and network requests that only appeared in the failing run.

Grouping duplicate failures across parallel CI runs

"We have 12 failed traces from this pipeline. Are they all the same failure?"

Call generate_error_signature on each — identical signatures mean identical root cause, no need to read every trace.

Diagnosing which API call caused a DOM change

"The modal appeared but I don't know which fetch triggered it."

correlate_dom_and_network joins the HAR log and DOM snapshots automatically. Example output:

{
  "total_correlations": 1,
  "correlations": [
    {
      "action_id": "4:Locator.click",
      "triggering_request_url": "https://api.example.com/cart/items",
      "response_status_code": 200,
      "response_body_snippet": "{\"items\":[{\"id\":\"abc\",\"qty\":1}]}",
      "time_to_dom_mutation_ms": 38,
      "resulting_dom_mutations": [
        { "type": "added", "selector": "heading \"Cart (1 item)\"" },
        { "type": "removed", "selector": "button \"Add to cart\" [disabled]" }
      ]
    }
  ]
}

Performance timeouts — not just missing elements

"The test times out on goto, but there's no JS error. What's blocking the page?"

detect_performance_anomalies inspects screencast-frame gaps and flags Long Tasks. Example output:

{
  "anomalies": [
    {
      "kind": "slow_action",
      "blocked_action_id": "2:Frame.goto",
      "task_duration_ms": 4200,
      "threshold_ms": 500,
      "concurrent_network_load": 9,
      "frame_drop_count": 0,
      "worst_frame_gap_ms": 0,
      "suspected_cause": "network_saturation"
    }
  ],
  "suspected_memory_leak_flag": false,
  "p50_action_duration_ms": 95,
  "p95_action_duration_ms": 780,
  "total_frame_drop_count": 0
}

suspected_cause distinguishes a blocked main thread (main_thread_blocked — frame gaps present), a waterfall of concurrent fetches (network_saturation — ≥5 in-flight), and a navigation/hard timeout (timeout_or_navigation — duration >3 s with no other signals).

Checking what Playwright version and HAR mode a trace uses

"The trace came from an unfamiliar CI configuration. Is the response body data available?"

extract_trace_metadata_strict inspects the archive before you run any other tool:

{
  "format_version": 6,
  "har_mode": "embed",
  "retry_sessions": [
    { "session_id": "s1", "failed": false },
    { "session_id": "s2", "failed": true }
  ],
  "failed_session_id": "s2"
}

har_mode: "embed" means body snippets are inline. "attach" means they're in separate resource files. "omit" means headers only — correlate_dom_and_network will return empty response_body_snippet in that case.