Set once. Vibe forever. An AI agent swarm that farms yield for you on Stellar, under limits you sign once and the chain enforces.
Live app: vibing-farmer.pages.dev · VF Wallet extension · Stellar testnet
Yield farming is the same loop: find a vault, approve, deposit, do it again for the next protocol. Vibing Farmer turns that into one wallet signature. An AI strategist picks vaults and writes per-agent instructions; workers deposit in parallel. Network fee sponsored by fee-bump relay.
The AI does not get custody of your funds. Each agent runs in a disposable on-chain account with hard limits: how much it can deposit, which vault, until when. Those limits live in contracts (allowance, expiry, vault pin), not in a prompt.
- Open vibing-farmer.pages.dev.
- Create a VF Wallet (passkey-based, no seed phrase, no extension required). Freighter, xBull, and Albedo work on testnet if you prefer those.
- Get test USDC from VF Wallet's built-in faucet.
- Go to Strategy, set amount, risk, and number of agents, review the plan, then sign once.
- Watch agents deposit in parallel and follow decisions on the Agent dashboard. Network fee sponsored by fee-bump relay.
Everything runs on Stellar testnet. No real funds.
-
Strategy. You set deposit amount, risk, and vault count. The AI returns an allocation plan and a skill file per agent, using live DeFiLlama data. A Monte Carlo pass stress-tests the allocation over 200 scenarios before anything runs.
-
AI council. Three specialists (yield, risk, market) score the proposal on their own. Disagreements go to a synthesis round. Verdict, cited playbook rules, and conflict resolution are logged for review.
-
Review. Skill files open in the Skills Drawer. Edit caps, expiries, or targets. Nothing runs until you approve.
-
One signature. You sign
funding_router.grant(budget + permission lifetime). One selected permission lifetime is encoded as an agent Unix expiry and a SEP-41 allowance ledger cutoff derived from the same captured start time. The allowance is the leash: the router deploys a fresh, scopedagent_accountper worker and can only pull what you approved. -
Parallel deposit. Workers sign with ephemeral ed25519 session keys. A fee-bump relayer sponsors each transaction. One worker failing does not abort the others. Network fee sponsored by fee-bump relay. (A Base pool in the plan settles as a sibling leg — see Optional Base leg for its extra prompts.)
-
Attestation. The strategy JSON is hashed and written on-chain so anyone with the original file can check what was approved.
-
Autonomy. A monitor loop polls positions, flags APY drift, and can propose rebalances. Each cycle goes back through the council. A keeper compounds on a cron; lifeboat radar can de-risk the vault at ledger speed under a user-signed mandate.
-
Kill switch. Two exits you can sign yourself, even if every server is down:
- Global:
token.approve(router, 0)— zero the allowance and funding stops. - Per agent:
agent_account.revoke()— flips an on-chain flag that authorization checks fail closed on.
- Global:
A Base pool only appears in the plan when the cross-chain relayer answers healthy. It settles alongside the Stellar workers in step 5, but it uses a separate Base flow and costs extra prompts:
- First Base run: worst case 4 prompts — grant + passkey setup (once, ever) + wallet-signed CCTP approve + burn.
- Every run after: 3 wallet signatures (grant + approve + burn) plus a passkey login confirmation. The login never goes away.
Withdraw that position anytime from the dashboard.
Scope is enforced by contracts, not promises:
- Agent accounts are deposit-only: pinned vault, amount cap per period, hard expiry. Approve, transfer, and anything else fails closed.
- The router holds no funds and has no admin or upgrade path.
- Vault hardening includes a share-inflation guard, untrusted-strategy NAV clamps, balance-delta verification, and emergency de-risk / quarantine hatches.
- The fee-bump relayer only sponsors an allowlisted set of operations. Both kill switches work without it.
Threat model, verified controls, test evidence, and residual risks: SECURITY.md. Testnet software; not an independent audit.
User input (amount, risk level, vault count)
|
v
AI strategist (Venice AI — key or x402; DeepSeek proxy fallback)
|-- Multi-vault allocation + live DeFiLlama data
|-- Skill JSON per agent (deposit cap + expiry)
|
v
Monte Carlo sim (200 runs, 30d)
AI Council (yield + risk + market specialists)
|
v
User connects wallet + single-signature grant (funding_router)
|
v
OrchestratorAgent --- attest strategy hash on-chain
|
+-----+-----+
v v v
Worker Worker Worker (parallel agents)
ed25519 session key signs a Soroban auth entry
Network fee sponsored by fee-bump relay.
autofarm vault → Blend Capital v2 (real testnet lending yield)
|
v
Autonomous monitor loop + keeper compound + lifeboat radar
Primary chain: Stellar / Soroban. Optional cross-chain leg to Base via Circle CCTP v2 + ZeroDev session keys.
The Explorer tracks 7 Soroban source crates, 6 first-party Vibing Farmer deployments, and 2 external protocol contracts: 8 static Stellar testnet addresses in total. Agent accounts are created dynamically per run and are not static deployments.
| Ownership | Contract | Address |
|---|---|---|
| first-party | Funding Router V2 (current app) | CB675TTSFM6COTGHGB7K2I7IODPQ3HTHOTTTXU2LJHXXNGTS45NOTRSE |
| first-party | Autofarm vault (live deposit, vfVLT 7-dp) |
CDWHNHIHOGBPXAK23NCU37BCXRRHCNNCEG6IPE4Q7FXBYLTJ7UYYKM77 |
| first-party | Blend strategy | CAR7XFFRKMUYSERYBSLQ4LXRY2E2W7G7WG4VQI55FWLSJWQVLNTAFVBE |
| first-party | Exit router | CDGDIPHBN3MSNURDX33IZBXXQTJPT7THAXSMVBAIOIXLOA6OF32IRS2J |
| first-party | Strategy attestation | CDDOW2FZ7ALBWBXF22TPMPDHPXSKTMLQGGQWUYX7YOJZAHICD7DUO2K6 |
| first-party | Agent registry | CAP5E2FPDAGEQ7SR55YRY4Z56GPBSTRRZJCYN2PQ6PZQHQJKYEDVM5FB |
| external | Blend v2 pool | CCEBVDYM32YNYCVNRXQKDFFPISJJCV557CDZEIRBEE4NCV4KHPQ44HGF |
| external | Stellar testnet USDC token (7-dp) | CAQCFVLOBK5GIULPNZRGATJJMIZL5BSP7X5YJVMGCPTUEPFM4AVSRCJU |
Look up any address on Stellar Expert:
https://stellar.expert/explorer/testnet/contract/<address>. The manifest also records WASM
hashes and deployment receipts; an agent WASM value is a hash, not a contract address:
deployments/stellar-testnet.json. The retired V1 router
remains a relay compatibility address in that manifest, but is not the active app router or an
Explorer static-deployment count.
| Layer | Technology |
|---|---|
| Smart contracts | Rust, Soroban SDK, OpenZeppelin Stellar contracts |
| Frontend | React 18, Vite 5, React Router v6, Framer Motion, react-force-graph-2d |
| Chain client | @stellar/stellar-sdk, Stellar Wallets Kit (Freighter / xBull / Albedo) |
| Wallet | VF Wallet (passkey smart account + extension) or any standard Stellar wallet |
| AI | Venice AI via API key or x402 (SIWE, prepaid USDC); DeepSeek server proxy as zero-config fallback |
| Yield | Autofarm vault → Blend Capital v2 (testnet lending interest) |
| Live market data | DeFiLlama API (APY, TVL, 7-day history); Tavily search for strategy context |
| Network fee | Sponsored by fee-bump relay (/api/stellar-relay, allowlisted ops) |
| Cross-chain (optional) | Circle CCTP v2 + relayer + ZeroDev on Base Sepolia |
| Crypto | ed25519 session keys; libsodium KDF-sealed per-worker key vault |
| Hosting | Cloudflare Pages: static SPA + /api/* Pages Functions |
| Route | Description |
|---|---|
/ |
Landing; no wallet required |
/home |
Portfolio, positions, alerts, market pulse |
/strategy |
Wizard: input → connect → skills → permission → execute → done |
/agent |
Dashboard: scopes, revoke, monitor status, journal, decision log |
/history |
Tx and strategy history |
/settings |
Wallet, permissions, agent config, language, skill source |
/explorer |
Stellar deployment facts and on-chain attestations; no wallet |
/replay |
Timeline replay from static JSON (no RPC) |
One typed skill file per agent (deposit-only; amounts in 7-dp base units):
{
"agentId": "worker-agent-1",
"vaultAddress": "CDWHNHIHOGBPXAK23NCU37BCXRRHCNNCEG6IPE4Q7FXBYLTJ7UYYKM77",
"skills": {
"deposit": { "maxAmount": "1000000000", "vaultAddress": "CDWHNHIH…KM77", "expiresAt": 1749686400 }
},
"generatedBy": "venice-ai",
"approvedByUser": true
}maxAmount is 7-dp base units (1000000000 = 100 USDC). Every field is editable in the Skills Drawer before approval. You can load custom skill files in Settings.
cd frontend
cp .env.example .env.local # Vite client vars
cp .dev.vars.example .dev.vars # server proxy + relayer secrets (Pages Functions)
npm install
npm run dev # http://localhost:5173AI keys are optional. Paste a Venice key in Settings, set a server-side DEEPSEEK_API_KEY, or use neither: a deterministic fallback keeps the demo working.
Server-side only (Cloudflare Pages env / .dev.vars — never VITE_ for secrets):
STELLAR_RELAYER_SECRET=S... # fee-bump sponsor (fund on testnet)
SOROBAN_RPC_URL=https://soroban-testnet.stellar.org
STELLAR_NETWORK_PASSPHRASE=Test SDF Network ; September 2015
SOROBAN_VAULT_ADDRESS=CDWHNHIH…KM77 # autofarm vault
SOROBAN_ROUTER_ADDRESSES=CB675TTS…TRSE,CCEWWRQV…CYE5 # canonical V2,V1 order
SOROBAN_AGENT_WASM_HASHES=1fdbe175…fbe1,d61ceaaa…a2ba # matching ordered hashes
SOROBAN_ROUTER_ADDRESS=CB675TTS…TRSE # singular current-V2 compatibility fallback
ALLOWED_ORIGIN=https://your-project.pages.dev # /api/* origin allowlist
DEEPSEEK_API_KEY=sk-... # optional AI fallback (BYOK-first)
TAVILY_API_KEY=tvly-... # optional market searchLeave host AI keys unset for a lockdown deploy (users bring their own keys).
Contract commands are WSL-only on Windows. From a WSL shell, change to this checkout's
soroban/ directory (for example /mnt/<drive>/<path-to-repo>/soroban), then run:
cd /mnt/<drive>/<path-to-repo>/soroban
stellar contract build
cargo test # unit + integration + security drills
cargo clippy --all-targets -- -D warningsDeploy and seed scripts live in scripts/soroban/. Addresses land in deployments/stellar-testnet.json.
cd frontend
npm test # Vitest suite
npm run lint # ESLint flat config (all warnings, no gate)
npm run lint:ci # ESLint gated against the checked-in warning-fingerprint baseline
npm run lint:warnings:update # regenerate the baseline locally after a reviewed warning change (refuses under CI=true)
npm run build # production → dist/
npm run build:ext # VF Wallet extension → extension-dist/
npm run pages:dev # build + wrangler pages dev (Functions locally).github/workflows/frontend.yml triggers on every push to main/dev, every pull_request, and merge_group (no narrowing types/paths filters — the gate always evaluates). Six stable jobs always report a result: frontend-unit-build (npm ci, lint:ci, format:check, brand:check, unit tests, build, build:ext, manifest:check, banned-string scan), relayer, keeper, soroban (pinned Ubuntu 24.04 + Rust 1.93.0 + stellar-cli 26.1.0 — the CLI's MSRV, which satisfies the SDK's 1.91.0 MSRV, not the plan's stated 1.82.0 — stellar contract build, cargo test --locked, cargo clippy --locked --all-targets -- -D warnings), playwright (test:visual, uploads the report/traces on failure), and claim-evidence (public-claim scan, evidence matrix, and feature-freeze validation). release-gate (if: always(), needs on all six, evaluated by scripts/ci/release-gate.mjs) is the intended single required check — it fails if any job is anything other than success, including skipped. The checked-in candidate forbids production publishing; if a GitHub Release event reaches claim-evidence, it fails with policy status 1 before the gate. deploy needs only release-gate, so no other job can be bypassed to reach Cloudflare Pages: dev → preview, main → the production GitHub Environment, with a non-secret readiness check before the traffic-shifting wrangler pages deploy step and serialized (non-cancelling) concurrency per ref.
Two of those guarantees are repo settings, not YAML, and are not yet in place: release-gate still has to be registered as the required status check in branch protection, and the production environment has to be created and protected — frontend.yml only names it. The playwright job is also merge-blocking on 47 visual baselines that were frozen on a developer box and have never run on a runner, at zero pixel tolerance. See GETTING_STARTED.md section 8 for the full release-prerequisite list, including source-only V3/V4 activation, the three Base-recovery evidence blockers, and the relayer's production boot ordering (D1 0005/0006 before the relayer, never after).
soroban/contracts/ # funding_router, agent_account, autofarm_vault,
# blend_strategy, registry, attestation
frontend/src/stellar/ # Soroban client, session keys, relay client, wallet kit
frontend/src/strategy/ # decision engine: MDP, Monte Carlo, council, gates,
# monitor loop, decision log, playbook curator
frontend/api/ # Pages Functions: ai, search, stellar-relay, faucet, guard
frontend/extension/ # VF Wallet browser extension source
relayer/ # optional Node CCTP + Base/ZeroDev cross-chain relayer
keeper/ # compound cron Worker + lifeboat radar
deployments/ # live contract manifests (Stellar testnet, Base Sepolia)
| Document | Focus |
|---|---|
| prd.md | Product requirements and feature status |
| SECURITY.md | Threat model, verified controls, residual risks |
| GETTING_STARTED.md | Local setup and demo checklist |
| EVIDENCE_MATRIX.md | Release claim evidence and feature-freeze policy |
| DESIGN.md | Design system / UI |
| soroban/README.md | Contract build and test |
- Stellar Developers · Soroban smart contracts
- Blend Capital · Stellar Wallets Kit
- Circle CCTP · ZeroDev
- Venice AI · DeFiLlama API · Cloudflare Pages
MIT