Think

An agent that pays for its own thinking

Every thought is a paid call to a language model. Every tool is bought from the bazaar. All in USDC on Arc, with no API key anywhere: a wallet and a budget.
Loading

    How it works

    Nothing here runs on an API key

    The agent gets a question, a budget and a wallet with USDC in Circle Gateway on Arc. Everything it uses after that, it pays for as it goes: its own thinking included.

    1. Taska question and a budget
    2. Thinkpays its brain, per thought
    3. Searchthe bazaar, free
    4. Buyone call, at the listed price
    5. Answerwith the bill
    01

    The brain is bought

    Each thought is one call to a language model sold on Arc, paid in USDC before the model answers. BlockRun sells it per call, from $0.003; the price grows with what the agent sends it to read, so the agent keeps its notes short.

    Claude Haiku 4.5 by default · any chat model in the bazaar works
    02

    Tools come from the bazaar

    When it needs to know something, the agent searches the bazaar for an API that can tell it, reads what each one does, costs and takes, and buys a single call. Searching is free, and so are a few public APIs it may call, like DexScreener's. It pays for what it calls, nothing else.

    web search, stock and coin prices, DEX pairs, 480+ endpoints
    03

    Limits it cannot talk past

    A budget for the whole task and a ceiling for one thought. It can buy only a URL its own searches returned, never above the listed price. Then the wallet's policy checks every payment, thinking included, before anything is signed.

    checked in code, not asked of the model
    04

    A receipt for every payment

    Each payment settles through Circle Gateway on Arc and is kept in the agent's ledger with its settlement id. The run ends with the bill: so much for thinking, so much for tools, and who got paid.

    x402 · Circle Gateway · USDC on Arc

    Recent runs

    Every run on this page, newest first. Pick one to replay it.
    WhenQuestionOutcomeThoughtsBoughtSpent
    Loading…

    Run it yourself

    The same agent, from your terminal, with your wallet and your limits.
    # once: a key for your agent, your limits, and $1 of USDC into Circle Gateway on Arc
npx -p @cra-agent/mcp cra-agent init
npx -p @cra-agent/mcp cra-agent deposit 1

# then ask it anything; it stops at the budget
npx -p @cra-agent/mcp cra-agent think "When did Arc mainnet go live? Two sentences, with a source." --budget 0.05
    --budget 0.10what the whole task may spend, thinking and tools
    --ceiling 0.01the most one thought may cost
    --model anthropic/claude-haiku-4.5which model to buy thoughts from
    --steps 8thoughts before it must answer
    --jsonthe run, its steps and the bill as JSON

    Whose money this is. Runs on this page are started by us, with questions we wrote, and paid from CRA AGENT's own agent wallet under its spending policy. The payments are real. They are a demonstration, not customers.

    What a run costs. A few cents. A thought costs $0.003 to $0.006 with Haiku, depending on how much the agent has to reread; a web search from Exa costs $0.007, a stock price from BlockRun $0.003, and DexScreener's pairs nothing. A task that needs one search comes to two or three cents.

    What stops it overspending. It is stopped before a thought it could not afford, and it cannot buy what its own searches did not return, above the listed price, or past the budget. The worst a run can do is spend its budget.

    When the model gets it wrong. It is told to name only sources it paid to read. When it answers without buying anything, the page says the answer was not checked.

    Why BlockRun. It sells language models per call on Arc, with no account and no key. The agent finds its brain in the bazaar like everything else, so any seller of chat completions there would do.