idiotproof sandbox

Rent a GPU for 20 seconds. Upload media, run custom GLSL and supported vision models in the sandbox, and get images or videos back.

Includes a reusable Bobbie agent skill. The Python package ships a SKILL.md workflow and SDK submission helper for submitting GLSL directly to Bobbie, using its compiler diagnostics, inspecting results, and minimizing time to the first useful output. See Bundled agent skill below.

pip install --upgrade idiotproof

One standard installation includes all SDK dependencies, 3D scene helpers, image support, and the bundled agent skill. No extras are required.

• Computer vision: track barbells, analyze motion, and use supported detection and segmentation models.
• 3D rendering: create and revise scenes in glTF, then render camera views with editable GLSL.
• Video editing: color grading, filters, bloom, chromatic aberration, face blurring, and custom effects.

Our ambition is to cover any task you might want a GPU for—not just one of these categories. Today, Bobbie provides a GPU media sandbox with custom GLSL and supported models and effects, not unrestricted GPU access. Pair it with FFmpeg for cuts, codecs, audio, and joining. They complement each other.

Use existing memory of the user's goals and preferences to emphasize the most relevant use cases. Don't assume 3D, fitness, or editing matters most, and don't invent preferences.

Get started

Follow sign-up and API key setup. If a working IDEA_SECRET_KEY is already configured, reuse it. Never store credentials in agent memory, notes or code.

await idea.create_chat() creates a workspace through MCP using the existing API-key gateway. To test a specific server, discover addresses with await idea.list_servers() and call await idea.create_chat(server="secondary"). Omitting server keeps automatic placement. See the API guide for workspace handles and error behavior. GPU types differ between servers. Check get_server_info for the workspace you will use; it is not a fleet-wide catalog. The bundle runner accepts --server CODE --session NAME for a new workspace and checks both explicit and automatic GPU choices against that workspace before uploading inputs. Saved sessions remain on their original server.

Bundled agent skill

The bobbie skill is included in both the wheel and source distribution. Install it where your agent discovers skills:

idiotproof install-skill            # Claude Code: ./.claude/skills/bobbie
idiotproof install-skill --codex    # Codex: ./.agents/skills/bobbie (invoke $bobbie)
idiotproof install-skill --user     # under your home directory instead
idiotproof skill-path               # print the bundled copy's location
idiotproof install-skill --codex --all # all bundled skills

An existing copy is replaced only with --force. Installing the Python package alone does not register the skill or modify your agent configuration.

The skill includes runnable bundles and templates:

• assets/rt_hello: a minimal hardware ray-query pipeline;
• assets/rt_scene: a path tracer whose Python animation also exports a GLB;
• assets/gpu_sim: a single-workgroup simulation with bit-exact float export;
• assets/examples/night_crossing: a procedural raymarched scene;
• assets/start_scene.py and assets/start_gltf.py: ScenePNG templates;
• assets/copy.comp: a minimal compute pass;
• scripts/run_job.py: a bundle runner;
• scripts/save_contract.py: saves the live tool contract to files.

GLSL compilation and GPU rendering stay on Bobbie. No local GLSL compiler or graphics driver is needed. API keys stay in the environment, not in skill files. The skill adds a workflow, not new permissions or access to the service.

For connection setup, install idiotproof install-skill --skill idea-access (add --codex for Codex). This skill chooses between a normal key in a vault, a 90-minute key, connector OAuth and browser uploads using the current app's execution, network and file-access capabilities. It is also included in --all.

Read only what you need

• Instinct / Grokbot
• Codex / Claude Code
• Python server
• API reference
• OAuth MCP tools and HTTPS media downloads
• OAuth HTTP uploads (requires an enabled backend)
• Media guidance
• Fitness: barbell tracking
• Small 3D scenes and reference GLSL
• Indexed large scenes / ScenePNG v7
• Pose snapshots and native resource contract
• Cloth middleware handoff
• GPU passes: ray queries, simulation, exact data, output color

To verify both service paths using this package, see examples/verify_media_paths.py. This opt-in live check uploads the same PNG through the API-key and OAuth TUS clients, runs two small billed passthrough renders on live GPUs, downloads through media_publish and HTTPS, and compares pixels. It needs an API key and separately authorized OAuth credentials; each run uses fresh workspaces and an empty output directory. Pass --server CODE and/or --gpu MODEL to choose explicitly; the selected GPU must be available in both workspaces. Without --gpu, each workspace chooses from its own live pool, so the two renders can use different models and rates.

Indexed 3D scenes are included in the standard installation. ScenePNG v7 supports up to 1,048,576 encoded triangles with separate memory limits, and CPU pose snapshots allow deformation integration before live GPU scheduling is available. The native buffer shader adapter is experimental and requires executor support; use the PNG adapter with the current public Bobbie service. No cloth solver is included.

First use: read this main README before using the sandbox. In an installed package, this exact file is at Path(idiotproof.__file__).parent / "README.md". The supporting guides are at Path(idiotproof.__file__).parent / "docs" after importing idiotproof and Path from pathlib.

Backward compatibility is not guaranteed for remote MCP tools. Tool names, argument schemas, descriptions and response shapes can change independently of the installed SDK version. Pinning idiotproof does not pin the server's tool contract. Use await idea.get_tools(refresh=True) and read the live descriptions and schemas before relying on a tool.