idiotproof sandbox

GPU video editing for agents.

Instinct / Grokbot · Codex / Claude Code · Python server

Start

Create an API key · Estimate costs

pip install idiotproof
export IDEA_SECRET_KEY=your_key

Install FFmpeg and ffprobe for video inspection, splitting, and concatenation. Interfaces and tool schemas may change. The default API is https://dev.67idea.com/v1; override it with IDEA_API_BASE_URL or Idea(base_url=...). Use a key for that environment. Never print your key or upload tokens.

Discover tools before uploading or rendering:

import asyncio
from idiotproof import Idea

async def main():
    async with Idea() as idea:
        tools = await idea.get_tools()
        print([tool["id"] for tool in tools])
        # Inspect the chosen tool's current docstring and schema before calling it.
        # Example, if present: print(tools.submit_bobbie_job.__doc__)

asyncio.run(main())

Run a workflow

1. Create a workspace: chat_id = await idea.create_chat(). Reuse this ID.
2. Upload: upload = await idea.upload_file(chat_id, "clip.mp4"). Up to 600 frames returns workspace_uri; longer videos split automatically and return ordered parts and workspace_uris. Images are also supported.
3. Call a discovered tool using its current schema. Start with a representative preview: submit, read feedback, revise, and inspect before scaling out. Tool calls are not automatically retried.
4. Publish the resulting workspace path: published = await idea.media_publish(chat_id, output_path).
5. Preview through your environment: localhost HTML for local coding agents, the host's media channel for Instinct/Grokbot, or your authenticated app UI for a custom server. Follow the matching guide above. Treat signed URLs as transient credentials; keep them out of logs and ordinary chat.
6. Save wanted results with await idea.download_media(chat_id, output_path, "final.mp4"), or download_and_concat() for ordered parts. Server media expires.

For multiple assets, use upload_batch(); its nested results retain asset/part order. For errors, inspect IdeaError.code, status_code, and request_id. A 403 requires checking key permissions and environment before retrying.

Read locally when needed

• API reference: calls, return shapes, concurrency, downloads, and concatenation.
• Media guide: preview strategy, Bobbie pipelines, VLM grounding, and constraints. Read before processing media.
• Why idiotproof: motivation and use cases.

These files ship inside the package. Locate them with:

from pathlib import Path
import idiotproof

docs = Path(idiotproof.__file__).parent / "docs"