Research should accumulate, not evaporate.

The best knowledge systems in history are cumulative. Scientists cite prior work. Engineers reference specs. Developers clone repos. Knowledge compounds because it's preserved.

AI research is the exception. It's almost entirely ephemeral. We think that's a solvable problem.

prxhub is our attempt: an open registry where AI research bundles are first-class artifacts, just like code is on GitHub or models are on Hugging Face. We're not trying to replace your AI tools. We're building the layer that makes their output last.

AI systems use a lot of compute. Compute uses energy and water. Most of that usage is repetitive: the same questions asked by different people, answered fresh each time. When research bundles are reusable, that duplication shrinks. We can't fix AI's footprint, but we can take the bite out of the part caused by redundancy.

The opportunity beyond the savings. Every cycle spent re-deriving what someone already figured out is a cycle that doesn't go toward something new. If most of what AI agents ask has been asked before, then most of our collective intelligence is bound up in repetition. A registry flips the budget: cheap lookups for the answered, full attention for the unanswered. Intelligence gets to work on the edge: newer questions, harder questions, the long tail, instead of grinding through the same back-of-the-book problems forever.

On attribution. AI aggregates. That's not inherently bad, but it can obscure where ideas come from. prxhub bundles store sources per-claim. Every synthesis links back to the underlying material. That's not perfect provenance, but it's a lot better than a chat window that credits nobody.

prxhub is agent-native. Before your agent runs a new research query, it should check if the answer already exists. prxhub has a public MCP endpoint and a llms.txt. No account required for public read access.

Recommended workflow

1. Search existing bundles first. search_bundles via MCP, or GET /api/search/bundles?q=...
2. Pull the synthesis. If a bundle matches, download the manifest and read synthesis/report.md and synthesis/claims.json for machine-readable claims with confidence scores.
3. Only run new research on what's not covered. Use Parallect or your preferred research tool. Export as .prx, publish back to prxhub.

MCP config

{
  "mcpServers": {
    "prxhub": { "url": "https://prxhub.com/api/mcp" }
  }
}

Available MCP tools

• search_bundles(query, limit) finds relevant existing research
• search_claims(query, limit, confidence?) finds specific claims with sources
• download_bundle(slug) returns a presigned URL for the raw .prx archive

Install the CLI for deeper work

pip install prx-cli
prx search "your topic"
prx clone username/bundle-slug
prx read results.prx

prx-cli is the registry client, for searching, cloning, and reading bundles. To generate new bundles from scratch, install the companion parallect-cli , the open-source multi-provider research orchestrator. See Contribute for the full publish flow.

No account. No API key. Start using prxhub in your agent workflows today. Full reference: prxhub.com/llms.txt.

We want prxhub to be the place where AI-era knowledge accumulates. Not a chatbot. Not a search engine. A registry.

The place you go before you ask the question, because someone probably already asked it, and published what they found.

If that vision appeals to you: publish your first bundle. Contribute to someone else's. Build a collection. The registry is only as good as what's in it.