Introduction

The problem

AI agents are starting to hold money. An agent that earns fees over x402, runs a service, or manages a budget accumulates idle USDC — and idle stablecoins are a slow leak. On-chain lending markets pay real yield on that same USDC, but capturing it means a pile of things an agent (and its owner) shouldn't have to take on:

• Custody risk. To deploy capital, something needs signing authority over it. Hand that to a bot and you've recreated every "the bridge got drained" headline.
• Protocol complexity. Kamino, Marginfi, and the rest each have their own deposit/withdraw semantics, oracle dependencies, account models, and SDKs that drift with every release.
• Risk management. Chasing the highest headline APY is how you end up the exit liquidity in a 99%-utilized pool. Real yield management means caps, buffers, utilization haircuts, and knowing when not to move.
• Economics. On a small treasury, a rebalance can cost more in gas than the yield it captures. Naive automation churns capital into fees.

Olbos exists so an agent can hold a wallet and an intent — "keep my idle USDC earning, nothing risky" — and have the rest handled, without giving up control of the funds.

The one mental model

Think of it as hiring a treasurer for your agent.

• Your agent opens a treasury account it pays for per-use.
• The treasurer (the Olbos engine) can move money between approved savings venues — and nowhere else. It cannot withdraw to itself, send to a stranger, or change its own permissions. Those aren't promises; they're enforced on-chain.
• You (the owner) hold the master key. You see everything through a glass wall, and you can fire the treasurer instantly — with a signature, or by revoking its authority directly on the blockchain, which works even if Olbos itself disappears.

Everything in these docs is an elaboration of that picture.

Three layers, kept separate

Newcomers conflate these constantly, so pin them down once:

A strategy allocates across venues under a custody arrangement. They're independent axes. A "balanced" strategy might split capital between the Kamino and Marginfi venues, all held in a Swig smart account your wallet controls.

On the localnet demo you'll see venues named delphi and olympia — those are mock lending venues (Greek treasuries, on-brand) used for development. On mainnet the venues are kamino-main and marginfi. The mocks are deliberately not named after real protocols so a demo is never mistaken for real funds.

What Olbos is — and isn't

It is: a safe, autonomous idle-cash yield router for a single stable asset (USDC), deploying into established, independently audited lending venues (Kamino, Marginfi) within an owner-set risk policy, under non-custodial smart-account custody, with every action paid for and audited.

It is not (yet): a multi-asset portfolio manager, a forward-rate forecaster, a leverage or LP strategy engine, or a tax/MEV-optimizing system. It uses spot APYs and a greedy allocator with risk bounds — not mean-variance optimization. That's the right scope for a treasury: boring, legible, and hard to lose money with. The strategy roadmap is where richer behavior lives.

Who uses it, and how

• An agent (autonomous) integrates the SDK or runs the MCP server and manages its treasury conversationally or programmatically.
• An owner (human) watches through the dashboard and holds the break-glass.
• A self-hoster runs the whole open-source stack against their own RPC and keeps everything. (Open source isn't generosity here — a custody claim is only credible if the code requesting the authority is inspectable. See security.)

Next: How it works — the end-to-end lifecycle.