Skip to main content
Your agent transacts autonomously, but you set the guardrails. As an operator, you fund the wallet, define what the agent may buy, and configure the policies that govern every purchase decision — before the agent ever touches a payment.

​
Fund the wallet

Paid calls draw from your agent’s USDC wallet balance. You can add funds in two ways.

From the dashboard

Sign in at sellag.vercel.app and add funds via checkout. Card payments convert to USDC automatically.

Direct transfer

Send USDC to your agent’s wallet address on Base, Solana, or Stellar.
You can also start a funding checkout programmatically:
The response returns a checkout_url. Open it in a browser to complete the payment; the balance is credited to your agent’s wallet automatically once the card payment settles.

​
Just-in-time funding

If the wallet runs dry mid-session, a paid call fails with WALLET_UNFUNDED and your agent surfaces a funding link. Click it, top up, and the agent resumes — no need to restart the task.
Pre-fund a session budget before enabling autonomous buying. It’s smoother than reacting to empty-wallet interruptions, and your policies cap the downside.

​
Govern spend with SKILLS.md

SKILLS.md is your agent’s constitution — a markdown file that tells Sella how your agent is allowed to spend. It drives the Policy Engine, which gates every purchase before any payment is authorized.
SKILLS.md
Place this file in your agent’s working directory. Sella reads it at the start of each session and enforces it throughout.
Controls how much autonomy the agent has. Human-assisted requires you to approve each purchase and is the safe default. Autonomous buys within budget limits without prompting you. Read-only lets the agent discover and evaluate listings but never make a purchase.
A 0.0–1.0 dial that tunes exploration versus exploitation. Low values (closer to 0.0) favor proven, popular listings the agent already knows work well. High values (closer to 1.0) let the agent try novel or untested listings. Set it to match your risk appetite for the task.
Hard caps enforced per session and per purchase. The Policy Engine checks these limits before producing any x402 payment signature — an over-budget purchase never reaches the wallet, regardless of agent behavior.
Global schema requirements applied to every dataset purchase decision, on top of any per-search filters you specify. If a dataset is missing any field listed here, the Policy Engine blocks the purchase.

​
The Policy Engine

Every purchase attempt passes through layered governance: SKILLS.md gates, buytemp, commerce mode, purchase history, and contextual task-fit checks. A purchase must clear every layer to proceed. No single signal can override the stack.
The Policy Engine and SKILLS.md are actively expanding. Start with commerce_mode and budget — those are the highest-leverage controls available today.

​
Monitor usage

Track spend, purchases, and per-tool activity from Dashboard → Usage, or read it programmatically using the usage endpoint.
The response includes a breakdown of spend by session and by tool, giving you a clear picture of what your agent is buying and how much each task costs. To inspect the active policy configuration — the settings currently in effect for your agent — use the policies endpoint:
The response reflects the parsed state of your SKILLS.md: commerce mode, buytemp value, budget limits, and required fields. Use it to verify that your policy file was loaded correctly before a session begins.

​
Security recap

The following controls are always in effect, independent of your SKILLS.md configuration.

Keys never leave you

Private keys live in ~/.sella-wallet.json on your machine. Sella never has custody.

Budgets are enforced

Hard limits in SKILLS.md cap spend regardless of agent behavior.

Replay-safe payments

Each x402 payment carries a nonce and expiry, preventing replay attacks.

Rotate anytime

Re-authenticate to mint a fresh key and retire the old one.