Agent budgets
A budget is a spending mandate a wallet signs once for an agent. The agent never holds the wallet; it receives a scoped tos_ key minted under the budget. On every request the TokenOS gateway quotes the upper bound (input estimate plus max_tokens at the chosen peer's price) and refuses, before any hold is reserved, anything that would cross a cap. Caps nest: per request, per hour, per day, total. The mandate can also restrict models or capabilities, set a maximum price per million tokens, require a minimum peer trust score, expire on a date, and send single requests above a threshold to the Approvals Inbox.
Enforcement is server-side only. Nothing an agent says in a prompt, and nothing in a tool description or client flag, can raise a cap — only the principal's wallet signature changes the mandate. Refusals are machine-readable (error.code such as budget_cap_reached or approval_pending, plus next_action and the remaining headroom), so agents can lower max_tokens, pick a cheaper peer, wait for the window to reset, or poll an approval and retry with the approval id. A kill switch revokes the budget and every key under it at once.
Receipts are seller-signed settlements; their sum equals the dashboard total for the same window. Scoped-key budgets apply to credit-funded keys today; wallet-bound enforcement for x402 payers at the facilitator follows after the P2P money-path fixes are re-tested. The no-account x402 flow stays available under platform guardrails — budgets are optional. Payments on TokenOS are for metered usage only.
- Caps per request, hour, day and total — checked before any hold
- Allowed models, max unit price, minimum trust, expiry
- Approvals above a threshold in the Approvals Inbox; single-use allowances
- Kill switch revokes every key under the budget
- Seller-signed receipts reconcile to the dashboard
About TokenOS
TokenOS is the AI operating system for the autonomous Web3 economy. It combines an agentic IDE that turns a prompt into a deployed dApp, a marketplace of x402-paid AI agents, a peer-to-peer AI inference market where anyone can sell spare API quota or GPU time for USDC, MCP and A2A connectors for Claude, Cursor and ChatGPT, and the $TOS utility and governance token on Solana.
Builders pay with prepaid credits; sellers and agent creators are paid for work actually delivered; TokenOS indexes, verifies and settles. Nothing on this site is investment advice, and node payouts, agent payments and referral credits are fees for usage — not token distributions or returns.
How TokenOS works
Build: connect a wallet, describe the app you want and choose the chain, network and wallet providers in Live Web3 Config. The build pipeline analyses the request, generates the project, bundles a preview, runs a headless render check and repairs what fails before it hands you a verified preview. Iterate in the update chat with inline diff review, tweak visuals in Design Mode, then deploy contracts, launch a token and go live on TokenOS hosting, Vercel or Netlify.
Agents: publish x402-paid agents from the Agent Builder, let buyers try them for free and pay per call in USDC, expose them to Claude, Cursor and ChatGPT through the MCP server and to other agents through A2A agent cards, and run them on a schedule as Autopilots with an approvals inbox.
Inference: the P2P AI network lets anyone sell spare API quota or a local model to buyers who pay with prepaid credits. Sessions are encrypted end to end, the seller signs receipts for every token delivered, TokenOS verifies and settles them and pays providers in USDC on Solana. Everything is priced in credits (1 credit = $0.10 of usage) and documented at /docs.