three.ws Packages 3D Bodies, Agent Logic and Payments in One Open-Source Stack

AI agents are usually confined to chat windows. The team behind three.ws is building a broader stack: visible 3D characters, an agent runtime, on-chain identities and tools for distribution across the web, ChatGPT, Android and MCP. The project detailed its approach in the Hugging Face Blog, published August 29, 2026.

The open-source code is available under Apache-2.0, while the project’s models are hosted on Hugging Face. Its Forge interface can turn a prompt, one to six reference photos or a sketch into a GLB model without an account, API key or payment. The pipeline routes jobs across systems including Microsoft TRELLIS, Hunyuan3D 2.1, TripoSR and TripoSG-scribble, using free-first failover when services are unavailable or rate-limited.

Developers can choose Draft output at 12,000 polygons, Standard at 30,000 polygons with 2K textures, or High at 200,000 polygons with PBR and HD textures. A separate selfie workflow creates a rigged avatar in about five seconds on a CPU, using 468 facial landmarks, 52 ARKit blendshapes and 15 visemes.

The runtime is designed to make those models useful rather than merely decorative. It supports tool calls, typed memory, speech and animation, while canonicalizing roughly 20 bone-naming conventions so imported assets from platforms such as Blender, VRoid and Roblox can be retargeted without a fixed rig allowlist.

The project also uses x402, an HTTP 402-based payment standard, to let agents transact in USDC. That matters for builders experimenting with agent-to-agent services: compute, skills and API calls can become programmatic expenses instead of transactions requiring a person to approve every payment.

Source: Hugging Face Blog