diff --git a/scripts/token-measure-vs-playwright.cjs b/scripts/token-measure-vs-playwright.cjs
new file mode 100644
index 00000000..b4426e43
--- /dev/null
+++ b/scripts/token-measure-vs-playwright.cjs
@@ -0,0 +1,245 @@
+// Apples-to-apples token measurement: Provar MCP vs. Playwright MCP.
+//
+// Both servers expose tools/list via JSON-RPC stdio. This script drives each
+// server with identical methodology — initialize → tools/list — and reports
+// the catalog size (characters, approximate tokens at chars/4) plus a per-tool
+// breakdown for the heaviest items.
+//
+// For Playwright MCP we additionally measure a representative tools/call:
+// browser_snapshot on a sample page. That's the per-interaction cost that
+// dominates Playwright MCP's 114K-per-test figure.
+//
+//   node scripts/token-measure-vs-playwright.cjs
+
+'use strict';
+
+const { spawn } = require('child_process');
+const os = require('os');
+const path = require('path');
+
+const TMP = os.tmpdir();
+const PROVAR_ENTRY = path.resolve(__dirname, '..', 'bin', 'mcp-start.js');
+
+// ── Generic JSON-RPC stdio driver ───────────────────────────────────────────
+
+function driveServer(name, command, args, env, onConnect) {
+  return new Promise((resolve, reject) => {
+    const server = spawn(command, args, {
+      stdio: ['pipe', 'pipe', 'pipe'],
+      env: { ...process.env, ...env },
+      shell: process.platform === 'win32',
+    });
+
+    let nextId = 1;
+    const pending = new Map();
+    let buf = '';
+    let stderrBuf = '';
+
+    server.stderr.on('data', (chunk) => {
+      stderrBuf += chunk.toString('utf-8');
+    });
+
+    server.stdout.on('data', (chunk) => {
+      buf += chunk.toString('utf-8');
+      let nl;
+      while ((nl = buf.indexOf('\n')) !== -1) {
+        const line = buf.slice(0, nl).trim();
+        buf = buf.slice(nl + 1);
+        if (!line) continue;
+        try {
+          const msg = JSON.parse(line);
+          const cb = pending.get(msg.id);
+          if (cb) {
+            pending.delete(msg.id);
+            cb(msg);
+          }
+        } catch {
+          /* non-JSON output — ignore */
+        }
+      }
+    });
+
+    server.on('error', (err) => {
+      reject(new Error(`${name} spawn error: ${err.message}`));
+    });
+
+    const rpc = (method, params, timeoutMs = 30000) => {
+      const id = nextId++;
+      const req = JSON.stringify({ jsonrpc: '2.0', id, method, params }) + '\n';
+      return new Promise((rpcResolve, rpcReject) => {
+        pending.set(id, rpcResolve);
+        setTimeout(() => {
+          if (pending.has(id)) {
+            pending.delete(id);
+            rpcReject(new Error(`Timeout (${timeoutMs}ms) waiting for ${method} on ${name}`));
+          }
+        }, timeoutMs);
+        server.stdin.write(req);
+      });
+    };
+
+    (async () => {
+      try {
+        const result = await onConnect(rpc);
+        server.stdin.end();
+        // Allow a brief grace period for shutdown
+        setTimeout(() => server.kill(), 500);
+        resolve({ ...result, stderr: stderrBuf });
+      } catch (err) {
+        server.kill();
+        reject(err);
+      }
+    })();
+  });
+}
+
+// ── Helpers ─────────────────────────────────────────────────────────────────
+
+function tokenize(jsonValue) {
+  const s = JSON.stringify(jsonValue);
+  return {
+    chars: s.length,
+    tokens: Math.round(s.length / 4),
+  };
+}
+
+function reportCatalog(name, toolArr) {
+  const { chars, tokens } = tokenize(toolArr);
+  const perTool = toolArr.map((t) => {
+    const sz = tokenize(t);
+    return { name: t.name, ...sz, descChars: (t.description ?? '').length };
+  });
+  perTool.sort((a, b) => b.tokens - a.tokens);
+  return {
+    name,
+    toolCount: toolArr.length,
+    catalogChars: chars,
+    catalogTokens: tokens,
+    meanTokens: Math.round(tokens / Math.max(toolArr.length, 1)),
+    topTools: perTool.slice(0, 5),
+  };
+}
+
+// ── Provar MCP runner ───────────────────────────────────────────────────────
+
+async function measureProvar(label, env) {
+  return driveServer(
+    `Provar MCP [${label}]`,
+    process.execPath,
+    [PROVAR_ENTRY, 'mcp', 'start', '--allowed-paths', TMP, '--no-update-check'],
+    env,
+    async (rpc) => {
+      await rpc('initialize', {
+        protocolVersion: '2024-11-05',
+        capabilities: {},
+        clientInfo: { name: 'token-compare', version: '1.0.0' },
+      });
+      const tools = await rpc('tools/list', {});
+      return reportCatalog(`Provar MCP — ${label}`, tools.result?.tools ?? []);
+    }
+  );
+}
+
+// ── Playwright MCP runner ───────────────────────────────────────────────────
+
+async function measurePlaywright(label, extraArgs = []) {
+  return driveServer(`Playwright MCP [${label}]`, 'npx', ['-y', '@playwright/mcp', ...extraArgs], {}, async (rpc) => {
+    await rpc('initialize', {
+      protocolVersion: '2024-11-05',
+      capabilities: {},
+      clientInfo: { name: 'token-compare', version: '1.0.0' },
+    });
+    const tools = await rpc('tools/list', {}, 60000);
+    const report = reportCatalog(`Playwright MCP — ${label}`, tools.result?.tools ?? []);
+
+    // Try to measure a representative tools/call too — browser_snapshot
+    // against a simple page. This captures the per-interaction cost that
+    // Playwright MCP charges on every step.
+    try {
+      await rpc('tools/call', { name: 'browser_navigate', arguments: { url: 'https://example.com' } }, 60000);
+      const snap = await rpc('tools/call', { name: 'browser_snapshot', arguments: {} }, 60000);
+      report.snapshotTokens = tokenize(snap.result).tokens;
+      report.snapshotPage = 'example.com (simple page baseline)';
+    } catch (err) {
+      report.snapshotError = err.message;
+    }
+    return report;
+  });
+}
+
+// ── Output formatting ───────────────────────────────────────────────────────
+
+function fmtRow(s) {
+  return `${s.name.padEnd(58)} ${String(s.toolCount).padStart(5)}  ${String(s.catalogTokens).padStart(7)}`;
+}
+
+(async () => {
+  console.log('Apples-to-apples token measurement: Provar MCP vs. Playwright MCP\n');
+  console.log('Methodology: spawn each server, send initialize → tools/list, count chars,');
+  console.log('estimate tokens at ~4 chars/token. Numbers reflect what the MCP client');
+  console.log('serializes and sends to the LLM as its tool catalog.\n');
+
+  console.log('Measuring Provar MCP (3 configurations)...');
+  const provarStandard = await measureProvar('STANDARD (all groups, full descriptions)', {});
+  const provarCompact = await measureProvar('COMPACT (all groups, compact descriptions)', {
+    PROVAR_MCP_SCHEMA_MODE: 'compact',
+  });
+  const provarAuthoring = await measureProvar('AUTHORING (compact + inspect/connection/validation/authoring only)', {
+    PROVAR_MCP_SCHEMA_MODE: 'compact',
+    PROVAR_MCP_TOOLS: 'authoring,inspect,connection,validation',
+  });
+
+  console.log('Measuring Playwright MCP (default / out-of-the-box)...');
+  let pwDefault;
+  try {
+    pwDefault = await measurePlaywright('DEFAULT (out-of-the-box)');
+  } catch (err) {
+    console.error(`  ⚠ Playwright MCP measurement failed: ${err.message}`);
+    pwDefault = null;
+  }
+
+  console.log('\n══════════════════════════════════════════════════════════════════════════════════');
+  console.log(`Scenario                                                 Tools  ~Tokens`);
+  console.log('══════════════════════════════════════════════════════════════════════════════════');
+  console.log(fmtRow(provarStandard));
+  console.log(fmtRow(provarCompact));
+  console.log(fmtRow(provarAuthoring));
+  if (pwDefault) console.log(fmtRow(pwDefault));
+  console.log('══════════════════════════════════════════════════════════════════════════════════\n');
+
+  if (pwDefault) {
+    const ratioStd = (pwDefault.catalogTokens / provarStandard.catalogTokens).toFixed(2);
+    const ratioCpt = (pwDefault.catalogTokens / provarCompact.catalogTokens).toFixed(2);
+    const ratioAut = (pwDefault.catalogTokens / provarAuthoring.catalogTokens).toFixed(2);
+    console.log('Tool-catalog ratio (Playwright MCP / Provar MCP):');
+    console.log(`  vs Provar STANDARD :  ${ratioStd}× larger`);
+    console.log(`  vs Provar COMPACT  :  ${ratioCpt}× larger`);
+    console.log(`  vs Provar AUTHORING:  ${ratioAut}× larger\n`);
+
+    if (pwDefault.snapshotTokens) {
+      console.log(`Per-interaction cost (Playwright MCP — ${pwDefault.snapshotPage}):`);
+      console.log(`  browser_snapshot response: ~${pwDefault.snapshotTokens} tokens`);
+      console.log(`  (multiply by interactions per test to project the full session cost)`);
+    } else if (pwDefault.snapshotError) {
+      console.log(`Per-interaction measurement skipped: ${pwDefault.snapshotError}`);
+    }
+  }
+
+  console.log('\nTop 5 most expensive tools — Provar MCP STANDARD:');
+  for (const t of provarStandard.topTools) {
+    console.log(`  ${t.name.padEnd(42)} ~${String(t.tokens).padStart(5)} tokens  (desc: ${t.descChars} chars)`);
+  }
+
+  if (pwDefault) {
+    console.log('\nTop 5 most expensive tools — Playwright MCP DEFAULT:');
+    for (const t of pwDefault.topTools) {
+      console.log(`  ${t.name.padEnd(42)} ~${String(t.tokens).padStart(5)} tokens  (desc: ${t.descChars} chars)`);
+    }
+  }
+
+  process.exit(0);
+})().catch((err) => {
+  console.error('\nMeasurement error:', err.message);
+  if (err.stack) console.error(err.stack);
+  process.exit(1);
+});