Skip to main content

How to use Claude for crypto: what works, what doesn't, and how to set it up

Claude can read a 40-page protocol risk disclosure and tell you where the redemption gate is in under a minute. It cannot tell you where Bitcoin will be on Friday. Most disappointment with AI in crypto comes from asking for the second thing and skipping the first.

The useful split is not "trading versus research". It is what Claude can do from what you paste into it, what it can do once it has tools connected, and what it should not be asked to do regardless of tools. Since the Model Context Protocol (MCP) let assistants call external services, the middle category has grown quickly, and it now includes moving real funds under a permission you approve.

This page is the hands-on version of the question. It covers the tasks that work, the setup for connecting tools, the prompt patterns that get good answers, and the boundaries. The companion page, Can ChatGPT or Claude manage your crypto?, covers the custody model behind it.

What works with nothing connected​

Claude's strongest crypto use is reading. Paste a document and ask for what matters.

  • Risk disclosures and terms. Give it a protocol's docs or a fund's terms and ask: where is the redemption gate, what is the cooldown, who can pause the contract, what happens under negative funding. It reads the whole thing; you would skim.
  • Comparing two products on fixed dimensions. Paste both docs and ask for a table on custody, return source, exit path, fees, and access. This is how the comparison pages on this site start.
  • Explaining a mechanism. Ask how a funding rate is calculated, why a basis trade is hedged, or what a session key can and cannot do. It will explain at whatever depth you ask.
  • Reading code and transactions. Paste a contract function, an approval transaction, or a wallet's approval list and ask what it does. Claude is good at spotting an unlimited approval or an unusual permission.
  • Scenario arithmetic. Give it a position, a rate, a fee schedule, and ask for the net at three funding levels. It will not know the rate; it will do the maths on the one you give it.
  • Checklists and plans. Ask for a pre-deposit checklist for a specific product, or a plan to exit a position in steps.

The limitation in all of these is the input. Claude works from what you give it and its training. It does not know today's rates, a protocol's current state, or whether a document you pasted is the latest version. Paste the source, date it, and ask.

What works once tools are connected​

MCP lets Claude call a server that exposes tools. With the right servers connected, it stops guessing and starts reading live data or taking actions you have authorized.

Three kinds of servers matter for crypto:

  • Data servers expose onchain reads, prices, or exchange data. With one connected, "what is ETH funding on Hyperliquid right now" becomes a lookup rather than a refusal.
  • Analysis servers expose portfolio or risk tools that run on your addresses.
  • Action servers expose the ability to do something: deposit, delegate, withdraw. These are the ones where the permission model decides whether the setup is safe.

Deploy Finance runs one of the third kind at https://mcp.deploy.finance/mcp. Its tools split into two groups. Four are read-only and need no sign-in: what_is_deploy, how_yield_works, list_agents (live agents with APY, chain, and risk tier), and simulate_yield (a projection for an amount and an agent). The rest, covering deposits, delegation to an agent, and withdrawal, need a one-time OAuth sign-in to your Deploy account, and every fund-moving action hands off to the Deploy web app for you to sign with your own wallet. The signing and permissions guide describes that boundary.

What that gives you in practice: you can ask Claude to list the live agents, simulate a $5,000 deposit into Income: Funding Rates, explain the risks page, and then, if you decide to proceed, start the deposit from the conversation and approve it in the app. Claude never holds a key.

How to connect an MCP server​

The steps differ slightly by client, and the client documentation is the authority. The shape is the same.

  1. Pick the client. Claude Desktop, Claude Code, and claude.ai's connector settings can all connect to remote MCP servers. Claude Code adds servers with claude mcp add; the desktop and web clients have a connectors or integrations screen.
  2. Add the server URL. For Deploy, that is https://mcp.deploy.finance/mcp. No SDK install is needed for a remote server.
  3. Start read-only. Ask a question that only needs open tools. For Deploy, "what agents are live and what do they do" exercises list_agents without any sign-in. Confirm the tool call appears and the answer matches the docs.
  4. Sign in only when you intend to act. The OAuth sign-in grants the client scopes you approve. Read them.
  5. Approve every fund-moving action in the app, not in chat. If a flow asks you to paste a private key or seed phrase into the conversation, stop; that is not how any legitimate MCP server works.

Deploy's How to ask an AI assistant guide walks the read-only step, and Fund through MCP and Delegate through MCP cover the authorized steps.

Prompt patterns that get useful answers​

Claude answers the question you ask. In crypto, the difference between a vague and a precise question is the difference between a summary and a decision.

  • Give it the document, then ask for the failure mode. "Here are Protocol X's docs. Under what conditions can I not withdraw?" beats "Is Protocol X safe?"
  • Fix the dimensions before you ask for a comparison. "Compare A and B on custody, return source, exit path, and fees" gives a table you can act on.
  • Ask for the assumption, not just the number. "Simulate $10,000 in this agent for a year, and tell me what rate the simulation assumes." A simulation is an estimate; the assumption is what you are evaluating.
  • Ask it to argue against you. "I am about to deposit into this vault. Give me the three strongest reasons not to." It is better at this than at cheerleading.
  • Ask what it does not know. "What in this analysis depends on data you do not have?" It will tell you where a live source is needed.

Avoid asking it to be confident. Ask it to be specific.

What does not work, and will not​

Four requests fail regardless of setup. Knowing them saves the most time.

Price prediction. Claude has no information about future prices, and neither does anything else. A model that produces a forecast is producing a plausible sentence. Ask for scenarios with stated assumptions instead.

Trading signals. "Should I buy now" is a prediction with a verb. Claude can explain a setup, list what would have to be true for a trade to work, and compute the risk. It cannot tell you whether to take it.

Live data without tools. Unconnected, Claude does not know the current funding rate, gas price, or whether a protocol paused withdrawals an hour ago. It may answer as if it does. Connect a data server or paste the source.

Holding keys or signing. No assistant should hold a private key, and the correct architecture never asks it to. The action servers that work, including Deploy's, hand the signature to your wallet. If you find yourself pasting a seed phrase into a chat window, the design is wrong or the request is an attack.

Safety rules that do not bend​

The convenience of connected tools comes with two risks that are new to assistants and old to crypto.

Prompt injection. When Claude reads untrusted content, a webpage, a token description, a message someone sent you, that content can contain instructions aimed at the model. With action tools connected, an injected instruction could attempt a transaction. The defences are the ones Deploy's MCP enforces by design: read-only tools need no sign-in, fund-moving tools require OAuth scopes you approved, and every signature happens in the app under your own review. Do not connect action tools to an assistant that is also browsing untrusted pages in the same session.

Permission scope. Grant the narrowest scope that does the job, and revoke it when the job is done. On Deploy, an agent's session key can trade on whitelisted venues and cannot transfer or withdraw; the MCP sign-in is a separate, revocable grant. Review both on the permissions guide.

And the rule that predates AI: never paste a private key or seed phrase anywhere. Not into a chat, not into a form, not to a person claiming to be support.

A worked session​

Here is a realistic sequence, using Deploy's MCP as the example because its tools are documented.

  1. Paste the risks page and ask Claude to list every risk that applies to a market-neutral funding agent and rank them by how much they depend on Hyperliquid staying up.
  2. Ask it to call list_agents and describe the difference between Income: Funding Rates and Superstar in one paragraph each.
  3. Ask it to call simulate_yield for an amount and Income: Funding Rates, then to state the rate the projection assumes and what happens to the estimate if funding halves.
  4. Ask for the three strongest reasons not to proceed.
  5. If you proceed, start the deposit from the conversation and approve it in the Deploy app with your own wallet.

Every step where money could move was signed by you, in a wallet you control, after reading. That is the pattern to hold any AI-and-crypto workflow to.

Verdict​

Use Claude for reading, comparing, explaining, arithmetic, and planning, with the source pasted in. Connect MCP tools for live data and for actions you approve in your own wallet. Do not ask it for prices, signals, or anything that requires it to hold a key. Set up read-only first, sign in only when you intend to act, and keep action tools away from untrusted content.

Learn more​

Start with Deploy Finance​

Create a self-custodial Deploy Finance wallet and review the live agents.