Chibuzo360/heterotopia

A startup that runs itself on NEAR testnet. A founder creates a startup, investors buy tiered share NFTs, the team holds soulbound role NFTs, AI agents work for the holders under hard spending limits, customers pay for jobs through escrow, and revenue is split automatically into buckets that holders claim.
This is a labeled testnet simulation. No real funds are involved. DRYRUN=true is the default: the agent's "work" (a trading-signal text) is generated locally and is not real advice. Only the reference price comes from a live source (NEAR Intents 1Click dry: true quote, read-only).
Built at NEAR HACKATHON 2026. Licence: MIT.
## The problem
Giving an AI agent money normally means giving it a key that can do anything. Heterotopia gives agents a key that can only call two methods on one contract, and puts every other rule (caps, allowed payees, probation, kill switch) inside that contract, where the agent cannot argue with it.
## Live deployment (testnet)
| Item | Value |
| Contract | heto2-bybyte.testnet |
| Platform fee account | heto-fee-account.testnet |
| Demo agent | agent1.heto2-bybyte.testnet (function-call key only) |
| Network | NEAR testnet |
## Try it (no keys needed)
Each step is "do X, see Y". Replace RPC with https://rpc.testnet.near.org. Arguments are base64 JSON.
1. Do: run the contract's tests. cd contracts && cargo test. See: 159 tests pass (money split, escrow, refunds, agent limits, private results).
2. Do: run the agent library's tests. cd agents && npm install && npm test. See: 25 tests pass (look-alike payee guard, encryption round trip, quote parsing).
3. Do: read the agent's track record from the live contract:
curl -sS RPC -H 'content-type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"query","params":{"requesttype":"callfunction","finality":"final","accountid":"heto2-bybyte.testnet","methodname":"getagentreputation","argsbase64":"eyJhY2NvdW50X2lkIjoiYWdlbnQxLmhldG8yLWJ5Ynl0ZS50ZXN0bmV0In0K"}}'
See: a result byte array. It decodes to JSON with spendsok, spendsrefused, totalspent, timespaused, timesremoved. Refused spends are counted, not hidden.
4. Do: read a finished job (id 7) the same way, methodname getjob, argsbase64 eyJqb2JfaWQiOjd9Cg==. See: status Paid, and a result that starts with v1:. That is ciphertext. Anyone can read it on chain; only the customer's key opens it.
5. Do: read startup 6, methodname getstartup, argsbase64 eyJzdGFydHVwX2lkIjo2fQo=. See: the revenue buckets (shareholders, agent budget, operations) matching the split below.
6. Do: open the transactions listed under "Evidence" below in the NEAR testnet explorer. See: the real calls on chain.
### Run the full flow yourself (needs your own testnet funds)
The end-to-end script creates demo accounts, so it needs a funder account with about 10 testnet NEAR whose key is in ~/.near-credentials/testnet/. Deploy your own copy of the contract first (cd contracts && RUSTUPTOOLCHAIN=stable cargo near build non-reproducible-wasm, then deploy and call new), set FUNDERID and CONTRACTID, then:
cd agents
npm install
npm run setup # creates founder/investor/customer/vendor/agent sub-accounts, writes .env (git-ignored)