SmartTokenGuard

English · 简体中文

Stop burning credits on broken AI video.

Every black frame, frozen shot and off-model keyframe is money you already paid for. SmartTokenGuard checks each keyframe and clip from ComfyUI, Kling, Runway, Veo, Hailuo, Seedance or any other AI video tool the moment it renders, and tells your AI assistant whether it's worth paying for the next step: pass it, retry with exact fixes, switch models, or stop spending on that shot.

How much are broken renders costing you? Try the calculator: https://nwjnjjj.github.io/smart-token-guard/

Check the keyframe first. A rejected still costs nothing. A rejected 5-second clip has already been paid for.

How verdicts work: measured checks run automatically, and your assistant reviews the returned frames for hands, faces and identity before a render can PASS, so nothing broken slips through to the next paid step.

Plans

Checks reset on the 1st of each month (UTC). 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 Polar receipt is your API key. Cancel or change plans in the Polar customer portal. Run stg_pricing for the checkout links.

Install (MCP)

Requires uv.

Claude Desktop / Cursor: add to claude_desktop_config.json or ~/.cursor/mcp.json:

{
  "mcpServers": {
    "smart-token-guard": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/nwjnjjj/smart-token-guard", "smart-token-guard"],
      "env": { "STG_LICENSE_KEY": "your-license-key" }
    }
  }
}

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

VS Code: .vscode/mcp.json → {"servers": {"smart-token-guard": {"type": "stdio", "command": "uvx", "args": [...same...], "env": {...}}}}

Then ask: "Inspect renders/sh010_keyframe.png against refs/kai.png. If it passes, I'll render the clip on Kling (0.7 USD, fallback Runway)."

Python SDK

pip install git+https://github.com/nwjnjjj/smart-token-guard

from smart_token_guard import Client

stg = Client()  # reads STG_LICENSE_KEY
report = stg.inspect("renders/sh010.mp4", reference_image="refs/kai.png")
print(report["report"]["issues"], report["license"]["checks_remaining"])

decision = stg.qa_gate(model="Kling", render_cost_usd=0.7, technical_issues=report["report"]["issues"],
                       semantic_defects=["deformed_hands"], fallback_model="Runway")
print(decision["decision"], decision["actions"])

HTTP API: POST /v1/inspect (multipart file, optional reference, options JSON), POST /v1/qa-gate, GET /v1/account with Authorization: Bearer <license key>. See GET /v1/capabilities.

For AI agents (remote MCP and A2A)

No install needed. Media is passed as a public https URL; the server downloads it, analyses it and deletes it.

• Remote MCP (Streamable HTTP): https://stg-agent-560636228350.asia-east1.run.app/mcp with header Authorization: Bearer <license key>. Tools: stg_inspect_media_url, stg_qa_gate, stg_account, stg_pricing.
• A2A: Agent Card at https://stg-agent-560636228350.asia-east1.run.app/.well-known/agent-card.json, JSON-RPC message/send at /a2a. Send a data part {"media_url": "https://..."} (or text containing a link) to inspect, or {"skill": "qa-gate", ...} for a decision.
• Same plans and monthly checks as above. URLs must be public https hosts; private and internal addresses are refused.

Limits and privacy

• Up to 32 MB, 60 seconds and 1080p per file (JPEG, PNG, WebP, MP4, MOV, WebM, MKV). 30 requests per minute.
• Files are uploaded to the SmartTokenGuard API (Google Cloud), analysed in a temporary file, and deleted before the response is sent. Nothing is stored and nothing is used for training.
• The API receives your license key and the file. Project ledgers stay on your computer (~/.smart-token-guard/).
• Measured checks catch technical render failures. They do not certify anatomy, identity, copyright or story quality.

Support

Refunds, invoices and cancellation: Polar customer portal (link in your receipt). The client code in this repository is MIT licensed; the hosted API requires a paid license.