Venture 03

Asset Forge

A gateway that gives every generated asset a provenance.

Success

The idea

In development it can be hard to track the provenance of the different assets you use: which were AI-generated, and which were sourced from somewhere else.

The solution

Asset Forge is a strict MCP tool that gates every generative asset in my projects. Agents can't call OpenAI or ElevenLabs directly. They have to go through the gateway, which tags each asset with a deterministic UUIDv5 and records it in a ledger, so I know which assets are AI-generated and which came from elsewhere.

It also allows for iterative review. Assets are published in review batches, and my feedback is stamped to each asset's ID, so it survives regeneration.

The outcome

Success.

Every asset in my game projects has an ID and a ledger entry recording where it came from, and my reviews attach to that ID.

633ledger events
274assets with an ID
241AI-generated
129/180reviews approved
OpenAI122Gemini73AI, provider not recorded26ElevenLabs19Sourced (Freesound)17Derived from other assets12Procedural placeholder4ComfyUI (local)1AI-generatedSourced or proceduralDerived
OpenAI122Gemini73AI, provider not recorded26ElevenLabs19Sourced (Freesound)17Derived from other assets12Procedural placeholder4ComfyUI (local)1AI-generatedSourced or proceduralDerived
274 assets, each with its own ID, and the ledger knows which 241 were AI-generated. 129 of 180 asset reviews ended in approval. 62 prompt records are left out.Source: asset-forge/provenance/ledger.jsonl and feedback.jsonl; one asset per GUID, by its latest successful event.

The learnings

What went right

  • Observability of agents is important, but it needs to go deeper. If an agent is sourcing something from somewhere, making it do so through a gateway is a great tool.

What went wrong

  • Local asset generators consumed a lot of VRAM, so agents would spend time waiting for VRAM to free up, or for Asset Forge to come back online.
  • Even for simple requests, agents followed a rigorous plan that included review. In future I'd build a skill that tells the agent whether I want to review the assets, and control the context through skills rather than embedding it in the top-level context.