The model never holds the keys.
Agent vaults. Guardrails in the contract, not the prompt.
Rienda is a self-custodial vault for LLM trading agents on Robinhood Chain. The vault holds the capital and enforces the risk policy on-chain. The agent gets a session key that submits trade intents and does nothing else — it cannot withdraw, cannot change a limit, cannot grant itself either.
Losses get bounded, not prevented. That is the whole promise.
Nothing here accepts deposits today. The vault contracts deploy to Robinhood Chain testnet (chain id 46630) first; mainnet waits on an external audit and a legal review. What's already live is the rail underneath — see /security for what it does and doesn't guarantee.
Where the limits live.
Prompt-level guardrails fail the moment the model is jailbroken, hallucinating, or simply wrong. Rienda assumes all three will happen and puts the limits where the model can't reach them: the model proposes, the contract disposes.
Self-custodial
Deposits sit in a vault contract owned by you. VDM Nexus never holds a key and cannot move funds. The kill switch — halt trading, withdraw everything — sits above the agent's authority, where the agent can't reach it.
Guardrails in contract code
Position caps, loss limits, drawdown throttles, an asset whitelist. Every trade intent is checked against all ten before it executes. A jailbroken model still can't talk its way past a revert.
A decision history you can check
Each trade links to the Ed25519-signed receipt of the inference that proposed it. Anyone the owner shows it to can verify the chain — this exact model output produced this exact trade — without trusting us.
{{FOUNDER VOICE — replace with 3-5 sentences in your own voice. Structure: a specific moment + a specific dollar amount or trust failure + "that's when I started building." Reference: Nova Wallet's "$300K lost to malware" hook lands in 30 seconds. Avoid corporate framing. Avoid product features. Just the moment.}}
— Dennis van der Meulen, founder
Every intent passes all ten, or it reverts.
The guardrails compose — an intent has to clear each one to execute. None of them read the model's prose. They only read its orders, which is why none of them can be argued with.
Position caps + gross exposure
No single asset can exceed its configured share of the vault, and total open exposure is capped on top of that. An intent that would breach either reverts before it touches a venue.
Daily realized loss limit
Once realized losses hit the day's limit, the vault accepts risk-reducing intents only until the window resets. A bad day stops being a worse one.
Drawdown throttles
Past the first threshold, permitted sizing halves. Past the second, the vault goes risk-reducing-only. Recovery has to be earned at reduced size.
No leverage, structurally
The vault cannot borrow. No margin means no margin call, no liquidation cascade, and no path to owing more than the vault holds.
Agents that don't earn, don't think.
Inference costs money, so the vault meters it against performance. Each day the agent gets a survival minimum plus a share of the 7-day moving average of its realized PnL, hard-capped at the top. Unrealized gains buy nothing. Sustained losses shrink the allowance toward the minimum, and in that state the agent may only submit risk-reducing intents — enough compute to get smaller, not to dig deeper.
The allowance is spent through x402: per-call USDC, one signed receipt per call. The design assumes most trading strategies lose money. That assumption is why the budget exists.
The rail that makes the history checkable.
A vault bounds what an agent can lose. It doesn't, on its own, tell you why the agent did what it did. That part comes from the signed-inference rail VDM Nexus has been running on Solana and Base mainnet since 2026-05-21 — mainnet-live infrastructure, not a diagram.
Where it stands.
No dates. No performance claims — nothing has traded, so there is no track record to show yet.
Eight packages, still shipping.
Six on npm, two on PyPI, all MIT. Ed25519 identity, the x402 client, paywall middleware for Express / Hono / Next.js, an MCP server, and providers for the Vercel AI SDK, Mastra, and LangChain. They're what Rienda is built on, and they keep working standalone.
The rail is MIT. The vault contracts open with the testnet deploy.
Leave an email
One note when the VaultFactory hits Robinhood Chain testnet, plus the occasional build-log digest. If you're already shipping on the rail, say what you're building — that's the message worth reading.