For AI video creators · ComfyUI · Kling · Runway · Veo · Hailuo · Seedance

Stop paying for broken AI video renders.

SmartTokenGuard checks every keyframe and clip your AI assistant helps you make. It catches the failures you'd otherwise pay to re-render later, then tells the assistant what to do: pass it, retry with specific fixes, switch models, or stop spending on that shot.

Get a license — from US$9/month Install guide

What it catches

Blur & soft focus

Edge-sharpness measured the same way for 16:9, 9:16 and 4K, so dark or shallow-focus shots aren't falsely flagged.

Flicker & frozen frames

Real alternating flicker (not lightning or camera moves), stalled motion, clips that barely animate.

Black frames & exposure

Black or blown-out renders and over- or under-exposed keyframes, found before you pay for image-to-video.

Look drift

Colour and lighting compared against your reference image, so shot 40 still matches shot 1.

Hands, faces, identity

Review frames go to your assistant, which looks at them. A render only passes after that visual review.

Impacts, flashes, cuts

Exact frames and times, ready for placing sound effects in your editor.

How it saves money

Check the keyframe first. A rejected still costs nothing; a rejected 5-second clip has already been paid for.
Inspect the clip. Measured checks, plus review frames for your assistant to look at.
Follow the decision. PASSRETRY_SAME_MODELSWITCH_TO_FALLBACKHALTHALT_BUDGET
Retries come with the exact negative-prompt terms, a lower motion setting and a new seed.
See the spend. Cost per shot, attempts per passed shot, first-try pass rate, video renders avoided.

Pricing

Starter

US$9 / month
  • 300 checks per month
  • QA decisions and spend reports included
  • MCP server + Python SDK + HTTP API
  • Cancel any time
Starter monthly US$99 / year

Pro

US$29 / month
  • 3,000 checks per month (10x Starter)
  • For daily creators and studio artists
  • Everything in Starter
  • Cancel any time
Pro monthly US$290 / year

Checks reset on the 1st of each month. One license per person; teams buy one per member. Checkout, tax and invoices are handled by Polar, the merchant of record. The license key in your receipt is your API key.

Install in one minute

Requires uv. Claude Code:

claude mcp add smart-token-guard -e STG_LICENSE_KEY=your-license-key -- \
  uvx --from git+https://github.com/nwjnjjj/smart-token-guard smart-token-guard

Claude Desktop, Cursor, VS Code and Python SDK instructions are in the README.

from smart_token_guard import Client
report = Client().inspect("renders/sh010.mp4", reference_image="refs/kai.png")
print(report["report"]["issues"])

What it doesn't do

It never generates or re-renders video and never calls a paid AI API for you. It doesn't certify faces or anatomy by itself: your assistant judges those from the review frames. It doesn't keep your files. Uploads are analysed in a temporary file and deleted before the response is sent. Limits: 32 MB, 60 seconds, 1080p per file.