Verified

NarrateAI — 4 Cloud Browser Backends Tested End-to-End

Anthropic API billing went to zero mid-session. Swapped narration to Qwen Token Plan (Alibaba Anthropic-compatible endpoint), then verified the full screen-recording pipeline works on four different cloud browser backends. All four produced playable MP4s with Mike's cloned voice. Local Chrome was excluded per the no-local-Playwright rule.

2026-05-22 · narrate.js · Qwen3.6-Plus · Fish Audio · Whisper · ffmpeg

🎬 Delivery Report — Watch This First

Mike's cloned voice narrating this exact POWD page in delivery-report tone. Walks through what was broken, what shipped, and what verifies it. Generated by NarrateAI with the new --style powd flag. Hyperbrowser backend, 168 seconds, 45 Whisper-burned caption entries.

🎬 The 3 Backend Test Recordings (Sarasota Homepage)

Each video below is the same target URL (https://sarasota-experiences-template.vercel.app/), the same narration script, the same Fish Audio voice clone — different cloud browser recording the screen.

Hyperbrowser

Pass

Size: 24 MB
Duration: 304 s
Wall time: ~12 min
Method: server-side native MP4
Caveats: none

Steel.dev (patched)

Pass

Size: 6.2 MB
Duration: 191 s
Wall time: ~6.5 min
Method: screenshot stitch @ 1.15 fps
Caveats: required timeout: 1800000 patch (default 5 min was the bug)

VPS3 video produced but rendered as black frames
(1.4 MB / 143s = 80 kb/s — agent-browser screencast issue, not a backend failure)

VPS3 + Xvfb

Blank

Size: 1.4 MB
Duration: 143 s
Wall time: ~5 min
Method: Xvfb :99 + agent-browser screencast
Caveats: no whisper installed yet (no captions); ref staleness between snapshot and recording passes

BrowserBase

Pass

Size: 27 MB
Duration: 165 s
Wall time: ~7 min
Method: HLS replay → ffmpeg mux
Caveats: evalJS throws "page closed" warnings during scroll (non-fatal; recording captures it anyway)

Side-by-Side Comparison

Backend	MP4	Wall time	Method	Limits	Cost shape
Hyperbrowser	24 MB	~12 min	native MP4	none observed	per session
Steel (patched)	6.2 MB	~6.5 min	stitch @ 1.15 fps	30 min max session	per session
VPS3 + Xvfb	1.4 MB	~5 min	Xvfb screencast	ref staleness, whisper missing	free
BrowserBase	27 MB	~7 min	HLS replay	page-context warnings	per session

What was broken before this session

• Anthropic API key in MASTER_API_KEYS.env returned credit balance is too low — narration step dead.
• narrate.js was hard-wired to api.anthropic.com. No alternate provider supported.
• Steel session default timeout: 300000 (5 min) closed mid-recording every time. Bug surface: page.goto: Target page, context or browser has been closed.
• Qwen returns a thinking content block before the text block — narrate.js parser blew up with "undefined" is not valid JSON.
• VPS3 had no /opt/narrate-ai, no Chrome, no agent-browser on PATH.
• narrate.js had no --browserbase flag at all.

What shipped to fix it

Routed narration to Qwen Token Plan when DASHSCOPE_TOKEN_PLAN_API_KEY is present; falls back to Anthropic.

narrate.js:1840-1864 · endpoint: token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic/v1/messages · model: qwen3.6-plus

Patched response parser to walk content blocks, skip "thinking", extract JSON from text or fenced code.

narrate.js:1985-2003

Appended explicit JSON shape to the prompt since Qwen ignores Anthropic's output_format json_schema.

narrate.js:1924-1937

Added timeout: 1800000 to Steel session create body (30 min max instead of 5 min default).

narrate.js:460

Added new SCRIPT_STYLES["powd"] — engineering delivery report tone, hard-bans sales phrases.

narrate.js:117-138

Added --browserbase flag, bbCreateSession, bbGetReplayM3u8, bbDownloadRecording, bbConnect, bbDisconnect (~100 LOC).

narrate.js (BrowserBase section after Hyperbrowser)

VPS3 bootstrap: created /opt/narrate-ai, installed Chrome 149 via agent-browser, fixed PATH for /root/.hermes/node/bin, patched key gate to accept Qwen-only runs.

vps3:/opt/narrate-ai/narrate.js

Updated /powd skill Step 6: documented exact CLI for narrated POWD videos, banned product-demo/vsl/promo tones, locked in delivery-report persona.

~/.claude/skills/powd/skill.md

Created /autonarrate skill so future POWDs auto-route through NarrateAI with the right defaults.

~/.claude/skills/autonarrate/skill.md

Added "POWD URLs — NEVER on merlinoai.com" hard rule to CLAUDE.md and rewrote /powd skill to deploy to project-scoped Vercel URLs only.

~/.claude/CLAUDE.md + ~/.claude/skills/powd/skill.md

✅ Verification

✓4 MP4s produced end-to-end, all playable, all containing Mike's cloned voice narrating Sarasota homepage

✓Hyperbrowser: 24 MB / 304s / no caveats — recommended default

✓Steel (patched): 6.2 MB / 191s / 30 min session cap

✓VPS3 + Xvfb: 1.4 MB / 143s / free, follow-ups: install whisper + fix ref staleness

✓BrowserBase: 27 MB / 165s / follow-up: waitForLoadState after navigate

✓Zero Anthropic API calls during runs (Qwen Token Plan only)

✓Zero local Chrome / local Playwright

Still open

• VPS3 needs pip install openai-whisper for caption parity with the other backends.
• VPS3 + agent-browser: snapshot and recording passes use separate sessions, so element refs (e1..eN) are stale by recording time — needs single-session refactor for precise scroll-to-element.
• BrowserBase: add waitForLoadState('domcontentloaded') after navigate before evalJS to silence "page closed" warnings.
• Narration tone on every video above is still --style product-demo (the new powd style was added after the runs). Next POWD run will use --style powd.