Agent-driven memory leak auditing for websites, built on the
Chrome DevTools MCP server
and its memory-leak-debugging skill.
The DevTools MCP server already provides the engine: take_heapsnapshot,
interaction tools (click, navigate_page, fill), and a skill that encodes
the snapshot methodology (baseline → repeat interaction ~10× → post snapshot →
revert → analyze with memlab — never read
raw .heapsnapshot files). This project adds the orchestration around that
engine:
- Per-site test plans — analyze a URL, enumerate the interactions worth exercising (SPA route cycles, modals, infinite scroll, forms), and persist the plan so it's reviewable and re-runnable.
- A single-URL audit loop — run every scenario in the test plan through the snapshot methodology and emit structured, actionable findings (see schema/report.schema.json).
- A batch runner — point it at a list of URLs (e.g. HTTP Archive / Tranco top sites) and fan out headless audits.
- Aggregation — merge per-site reports into a study of the most common leak patterns across the web.
- A leak playground — a small site with deliberately seeded, labelled leaks. It's both the demo (watch the agent find and fix each one) and the eval suite (the audit loop is correct iff it finds everything in playground/expected-findings.json).
Layout
.claude/skills/memory-audit/ The orchestration skill (test plan → audit → report)
.mcp.json chrome-devtools-mcp config (isolated + memory tools)
playground/ Seeded leak demos, leaky vs fixed mode, ground truth
schema/ Findings/report JSON schema
runner/ Batch runner (headless `claude -p` per URL)
aggregate/ Merge reports into a cross-site summary
urls/ URL lists + notes on sourcing top-site lists
testplans/ Generated per-site test plans (committed, reviewable)
reports/ Audit output, one directory per site (gitignored)
Quickstart
# 1. Get the DevTools MCP server + its skills (one of): # a) Claude Code plugin (server + skills bundled): # /plugin install chrome-devtools-mcp@chrome-devtools-plugins # b) Or rely on this repo's .mcp.json (server only; the memory-leak-debugging # skill methodology is referenced from the memory-audit skill). # 2. memlab for snapshot analysis npm install -g memlab # 3. Run the playground npm run playground # serves playground/ on http://localhost:8080 # 4. Audit it — /memory-audit works inside Claude Code, Codex, Gemini CLI, # and Antigravity when run from this repo (see runner/README.md) /memory-audit http://localhost:8080 # 5. Batch mode — URLs as arguments or --urls <file> (defaults to # urls/sample.txt). Runs with Claude Code by default; also supports # --agent gemini | antigravity | codex (see runner/README.md for setup). # Reports land in reports/<agent>/<site>/ for cross-agent comparison. npm run batch -- https://example.com npm run batch -- --urls urls/sample.txt --concurrency 2 npm run batch -- --urls urls/sample.txt --agent gemini # 6. Aggregate findings across reports npm run aggregate
See PLAN.md for the roadmap and open questions.