Skip to content
Start a project

Kraterion by Nano Soft

Kraterion runs AI agents on storage you own, and records every run

Kraterion is a runtime for AI agents you can audit, on storage you own. Files sit on Walrus, owned by your Sui account, behind a standard S3 interface, and every agent run is recorded for replay. Nano Soft built the first version in about 6.5 weeks for Sui Overflow 2026, where it took 1st place in the Walrus track.

Platform
Web, on Sui and Walrus
Category
AI agent runtime and storage
Recognition
1st place, Walrus track, Sui Overflow 2026
Year
2026
Role
Product, design, build, launch
Kraterion run trace showing the question, each tool call, the cited answer and a replayable record

A recorded run: the question, each tool call and a cited answer.

Challenge

Storage and AI agents both ask for blind trust

Cloud storage asks you to trust the provider. Cancel and the files can go. Get locked out and they are gone. The provider can read what it stores, and leaving means paying to copy everything out.

AI agents add a second layer of trust. They reason over that data, the same input can give a different output, and the trace usually lives in a vendor's database. You cannot prove what an agent read or how a past answer was made.

The build had its own limits. We had about 6.5 weeks from kickoff to the Sui Overflow 2026 deadline. Walrus prices storage per object, up front, which does not match the per-gigabyte bill developers expect. And in July, on the way to mainnet, Sui switched off the public interface our chain layer used, so we moved it to the newer one.

Approach

A familiar surface with the chain underneath

We kept everything a developer touches familiar: the S3 API, Google sign-in and an OpenAI-compatible endpoint. The chain sits only where promises used to live. Users never handle crypto tokens; the platform pays storage and transaction fees from a shared reserve.

Ownership is enforced by cryptography, not by a privacy policy. Each file is encrypted before it is stored, and decryption keys are only released while your on-chain bucket allows it. Revoking access is one transaction, after which the platform cannot read your files.

To make decentralized storage bill like normal storage, we pooled it: each project gets one storage pool, uploads use capacity and deletes free it. Every agent run records its inputs, retrievals, tool calls and output, stored on Walrus and anchored on Sui. We are open about limits too: the on-chain package is marked unaudited, and roadmap items say coming soon.

  • Pool the storage, meter the pool

    One storage pool per project turns Walrus's per-object pricing into the per-gigabyte bill developers expect.

  • Never break the guarantee

    If a feature could only exist by weakening ownership, we looked for a different design.

  • Chain as the source of truth

    The database is only a fast mirror of on-chain state, kept in sync by a background service.

  • Plain words for verifiability

    The product talks about debugging, reproducing, auditing and owning, and keeps crypto language in the deep docs.

What we built

Storage, knowledge and agents in one runtime

Kraterion has three layers: storage you own, knowledge bases that cite their sources, and agents whose runs you can replay. Developers reach them through APIs they already use.

  • S3-compatible storage you own

    Standard S3 tools work unchanged; files are encrypted, stored on Walrus and owned by your Sui account.

  • Access you can revoke

    One on-chain transaction stops the platform decrypting your files, and cancelling the service does not delete them.

  • Knowledge bases with receipts

    Hybrid search over a bucket, and every answer carries hashes of the passages it used.

  • Agents with replayable runs

    Each run records the model, inputs, retrievals, tool calls and output, and can be replayed and traced.

  • MCP server and chat widget

    The same agents work in Claude Desktop and Cursor over MCP, or on any site with one script tag.

  • Per-agent credentials

    Each agent gets its own scoped, revocable identity; revoke it and its file and memory access stop together.

Outcome

1st place in the Walrus track, then mainnet

Kraterion took 1st place in the Walrus track at Sui Overflow 2026, announced by the Sui Foundation on 27 August 2026. The hackathon drew 747 projects from 58 countries, and Walrus was its most-entered track.

The Walrus Foundation featured Kraterion as a builder story and called the pooled-storage design "the decision worth copying". Kraterion has been live on Sui mainnet since 1 September 2026 and is in private beta.

Recognition
1st place, Walrus track, Sui Overflow 2026
Built in
About 6.5 weeks
Status
Private beta, live on Sui mainnet

Screens

Inside Kraterion

  • Kraterion home page: files in storage, indexed into knowledge, and an agent run that cites its source
    Files in storage, indexed into knowledge, then a recorded run that cites its source.
  • Kraterion agent settings with persistent memory switched on and its two memory tools
    Agent memory is opt-in, with two tools the agent chooses to use.
  • Kraterion documentation introduction page with sidebar navigation
    Documentation for running agents on storage you own.

Want a product like Kraterion?

A few lines are enough: what the product is, where it stands and when you need it. We reply to every message within 2 working days.

Email
hello@nano-soft.tech
Reply
Within 2 working days
Based in
Bucharest, Romania