The onchain marketplace for software and inference for agents, built on Robinhood Chain. Let Cursor's agent browse skills, prompts, MCP configs, workflows and code, read each scan report, buy with USDG over x402, and pay per call for model inference. No account, no API key. Payment is the authentication.
agentware-spend-safety:
- `buy_package` and `ask_model` spend real USDG from the agent wallet. Before
calling either one, tell the user what it costs and get explicit approval in
this conversation. Never spend unprompted.
- Quote first. Use `quote_purchase` before `buy_package` and `quote_inference`
before `ask_model`, and show the user the exact number. Quotes are free.
- The free tools (`agent_guide`, `list_packages`, `get_package`,
`list_models`, `quote_purchase`, `quote_inference`, `wallet_status`) never
move funds. Prefer them to answer questions, and only escalate to a paid call
when the user asked for the result.
- Read the scan before buying. Call `get_package` and report the scan verdict,
score and any issues. If the verdict is not clean, say so plainly and let the
user decide. A scan lowers risk. It is not a guarantee.
- Never run, install or load a purchased package without the user reviewing
it first. Show where it was saved and what kind it is. An `mcp` or `code`
package can execute on this machine.
- If a payment is refused for exceeding the listed price or `SPEND_CAP_USDG`,
surface the limit and stop. Do not retry, split the payment, or route around
it. Only the user raises the cap.
- If a paid tool fails for a missing `AGENT_KEY`, low USDG, or no ETH for the
one time Permit2 approval, run `wallet_status` and tell the user exactly what
to fund. Never ask the user to paste a private key into the chat.
- Keep `max_tokens` on `ask_model` close to what the answer needs. The price is
an upper bound that includes the full output budget.