SuperLibrary keeps what agents make and what work has been done. Files are uploaded as versioned artifacts, finished work arrives from SuperPipeline along with AgentPod links, and every item carries one of seven outcomes: approved, completed, in-progress, superseded, rejected, failed or abandoned. Agents can consult it over MCP before they start, and people browse it in a web app.
It exists because of a gap I hit in October 2026. SuperPipeline stores references, not content, and AgentPod only shows a workspace while it is running. So when my agents needed to publish their work, they put it in gists under my personal GitHub account, two of them public. I made them secret and took gh off the agent host, then started SuperLibrary so that “publish this” had a safe answer. The design doc names a second problem: an agent starting a card had no way to ask what had already been done, decided or tried.
How it works:
- It runs on Cloudflare Workers and D1, and serves uploaded content from a separate origin.
- Uploads are scanned for common key formats and size limits. The site is clear that this is pattern matching, not a guarantee.
- Full-text and vector search run together and rank by outcome: approved and completed work comes first, while rejected and failed work ranks lower but still shows.
- There is a
suliCLI and an MCP server with three read-only tools. Sign-in goes through the lab’s shared accounts service.
The first four phases (core, sources, retrieval and the library app) are live.