Using Claude to explain a crypto transaction

Claude is useful around a crypto transaction when the job is interpretation: explain a fee label, compare supplied fields, or turn a wallet prompt into questions. The model’s ability to call a tool does not give it the authority to approve the transaction.

Start with a supplied transaction

Give Claude a redacted quote or a transaction summary and state the boundaries:

Explain this transaction for a beginner. Use only the fields I supplied. Define asset, network, gas, slippage, minimum received, recipient, and digest. Mark missing values as unknown. End with the questions I should answer in the wallet before signing. Do not predict the outcome or send a transaction.

This prompt gives the model a source, a vocabulary target, and a review task. Ask it to show arithmetic when comparing two quotes. If an output changes after a new quote, retain both timestamps rather than blending them.

Understand Claude’s tool boundary

Anthropic’s tool-use documentation distinguishes client tools from server tools. Claude can return a structured call; for a client tool, the application executes it and sends back a result. A server tool can run on Anthropic’s infrastructure under its own contract. In both cases, the host and tool definition determine what is available.

That distinction is useful for a Sui workflow. A read-only tool can return public status or history. A review tool can prepare an app route. Neither result is a wallet signature. If a host exposes the SUI BROKERS participant kit, its six tools read public status, stage, wallet history, Desk, or partner data, and its action plan is explicitly review-only. It does not quote, build, sign, send, activate referrals, or mint an NFT.

Use three concrete review questions

For a quote, ask Claude to identify which values are quoted output and minimum received, whether the pair and amount match, and which expiry or route detail is missing. For a fee explanation, ask it to separate network gas from any application or partner commission. For a wallet prompt, ask it to produce a short checklist of recipient, assets, amount, network, and gas payment.

The reader then compares every answer with the current app and wallet. A model can misread a label or reuse stale context. If the wallet prompt differs from the summary, stop and resolve the difference before signing.

Read the result after confirmation

If you proceed, connect the intended wallet explicitly, review the final transaction, sign only after checking the prompt, and wait for confirmation. A copied digest is not a service record until the service reports its status and verification. An assistant can format that readback, but it should preserve pending, confirmed, partial, and unavailable as separate states.

The current Swap rule credits one Point per $4 of verified volume. A confirmed record with $120 of eligible volume corresponds to 30 Points under that rule. The example explains a ledger calculation; it does not say that a quote, prompt, or unconfirmed event will produce Points. The AI-assisted swap guide follows the same stages.

Check the host before relying on it

Claude Code documents MCP connections and scopes in its MCP guide. That documentation describes a host workflow, not a universal promise that every Claude surface exposes every server. Confirm the transport, endpoint, account permissions, and approval behavior for the host you actually use.

The practical role for Claude is clear: make the evidence easier to read, leave uncertain fields visible, and return the final decision to the person holding the wallet.