Connecting Claude to a crypto wallet means giving Claude an MCP server whose tools can read balances, build transactions or sign them. The Model Context Protocol is the open-source standard Claude uses to reach external systems, which the MCP project compares to a USB-C port for AI applications (MCP documentation). The setup choice decides what Claude can actually do with funds: read only, build unsigned transactions for a person to sign, or sign within a spending policy. This guide compares the three setups, shows how to install each in Claude Code and the Claude apps, and covers the limits to set before any key is reachable.
Last reviewed: Nov 2026. Tool names and flags are quoted from each project's documentation.
Three Ways to Connect Claude to a Wallet
There are three practical setups. A read-only data server lets Claude see prices and wallet balances but never move funds. An unsigned-transaction server lets Claude write and simulate a transaction that a separate wallet signs. A local server with wallet tools lets Claude sign within a policy. Each step up adds capability and adds risk.
Setup | Claude can | Claude cannot | Example | Sources |
Read-only data | Read prices, wallets, portfolios, DeFi positions | Sign, trade or withdraw | CoinStats MCP | |
Unsigned-transaction handoff | Compile, simulate and build transactions | Sign; a wallet outside Claude signs | Sauce MCP with --no-wallets | |
Local signing within a policy | Build, simulate and sign from a local wallet | Exceed the wallet's spending policy through those tools | Sauce MCP over stdio | |
Documentation only | Search integration docs while coding | Submit transfers or hold credentials | Eco docs MCP |
CoinStats is a clean example of the first setup. Its MCP server reads data across 100,000+ coins and 120+ blockchains and states plainly that it cannot move funds: no signing, no trading, no withdrawals (CoinStats MCP). For portfolio questions that is the right amount of access.
Read-Only vs Signing Access
Read-only access lets Claude answer questions about a wallet. Signing access lets Claude change the wallet. The two should be separate decisions, made per task, because a model that reads untrusted text, such as a token description or a web page, can be steered by it. A read-only server limits the damage of that steering to a wrong answer.
Signing access is where the risk changes category. Eco's guide to prompt injection attacks on agents that hold funds walks through how instructions hidden in fetched content can redirect an agent's tool calls. The practical rule is to start every wallet integration read-only or unsigned, and add signing only for a narrow task with a limit attached.
How to Connect the Sauce MCP Server to Claude
The Sauce MCP server, part of Eco's Sauce SDK, gives Claude tools to compile, simulate and build onchain transactions. It is started with the sauce mcp command, needs Node.js 24 or newer, and works with any coding agent that supports MCP, including Claude Code. Transaction-building tools return unsigned transactions; wallet tools are a separate, optional layer.
Eco's documentation lists what the server provides: building tools such as sauce_compile, sauce_simulate and sauce_tx; sauce_send to broadcast transactions a wallet has already signed; the Sauce skills; and, over stdio only, wallet tools (Eco docs: Sauce for AI agents). A server started with --no-wallets --no-stores lists 18 tools, and the default stdio server lists 27. No tool takes a private key or a password.
The steps:
Install the Sauce SDK and add the MCP server to the agent as described in Eco's installation guide.
Ask Claude to list its sauce_ tools to confirm the connection.
Give it a task and a chain, for example a USDC payment on Base, and ask it to compile and simulate only.
Review the program, the simulation result and the unsigned transaction before any signing step.
Simulation does not need a funded wallet: sauce_simulate accepts an assumeBalance value that gives the program's account a balance for that simulation only (Eco docs). The CLI builds and sends transactions on Ethereum, Base, Arc and Solana, plus testnets including Base Sepolia and Solana Devnet (Eco docs: Sauce introduction).
For documentation rather than transactions, Eco also runs a docs MCP server. In Claude Code it is one command, claude mcp add --transport http eco-docs https://docs.eco.com/mcp, and it provides documentation search only; it does not submit transfers or provide wallet credentials (Eco docs: Build with agents).
Claude Code vs the Claude apps
Claude Code adds local servers that run as a command, which is how the Sauce stdio server and its wallet tools are reached. The Claude apps connect to remote servers instead: Anthropic's help center says custom connectors using remote MCP are available on Claude, Cowork and Claude Desktop across Free, Pro, Max, Ultra, Team and Enterprise plans, added through "Add custom connector" (Claude Help Center). Eco's docs note that the Sauce HTTP server never signs and has no wallet tools, so a remote setup is a build-only setup by design (Eco docs). See Claude Code vs Cursor vs Codex for onchain development for how the coding agents differ.
Keep Keys Out of the Model: The Unsigned-Transaction Handoff
The unsigned-transaction handoff keeps every private key outside Claude's reach. Claude writes and simulates the transaction, the MCP server returns it unsigned, and a wallet the person already uses signs and broadcasts it. Claude never sees a key, and a person or a wallet's own policy approves each transaction.
Claude ---- sauce_compile / sauce_simulate ----> Sauce MCPClaude <--- unsigned transaction (sauce_tx) ---- Sauce MCP | vwallet you control (or a wallet's own MCP server) | signs, applies its own approval rules vbroadcast (in the wallet, or with sauce_send)
Eco documents three signing paths: sign the unsigned transactions in your own wallet; pass them to a wallet's own MCP server, which keeps its key away from the agent; or use the server's own wallets over stdio (Eco docs). Starting the server with sauce mcp --no-wallets disables the local wallet tools while keeping transaction building.
The docs are direct about the limit of that flag. It controls the MCP server's tools, not the rest of the machine. An agent running as the user can read wallet files and password sources, sign with other tools that apply no policy, or edit the policy file itself (Eco docs: Security considerations). Keeping the key on a different device or in a separate wallet service is what actually keeps it out of the model.
Set Limits Before You Connect
Set limits before a signing tool is reachable, not after the first transaction. The useful limits are a spending policy enforced by the wallet, an allowlist of tools Claude may call without asking, a human approval step for anything above a threshold, and conditions inside each transaction that make it revert if the outcome is wrong.
Wallet policy. Sauce's sauce_wallet_send simulates what it is given and holds it to the wallet's spending policy before sending (Eco docs).
Tool permissions. Claude's agent tooling supports declarative allow and deny rules plus permission modes (Claude Agent SDK: permissions). Allow read and build tools; leave signing tools requiring approval.
Human approval. See human approval for AI agent transactions for thresholds and patterns.
In-transaction conditions. A Sauce program can require that a balance or minimum output holds during execution, so the transaction reverts instead of completing at a bad state (Eco docs).
The AI agent security checklist collects these into a pre-funding review, and AI agent spend controls covers budget design.
Where This Falls Short
No MCP setup makes a signing agent safe by itself. Policies sit in files an agent running as the user may be able to edit, simulations can go stale before inclusion, and read-only servers still expose data that can be used to steer the model. Each setup narrows risk; none removes it.
Same-user access. Eco's docs state that an agent running as you can change a policy or delete the spending ledger (Eco docs).
Stale simulation. State can change between simulation and sending (Eco docs). Simulation tools compared covers the gap.
Chain coverage. Sauce builds and sends on four mainnets today; other chains need other tools.
Frequently Asked Questions
These are the common questions about connecting Claude to a crypto wallet.
Can Claude hold my private key?
It should not. With the unsigned-transaction handoff, Claude only ever sees unsigned transactions. Eco's Sauce MCP tools never take a private key or password (Eco docs).
Can Claude send crypto on its own?
Only if it is given a signing tool. With a read-only server like CoinStats MCP it cannot move funds (CoinStats).
Do I need a funded wallet to test?
No. Sauce compiles and simulates with an assumed balance (Eco docs).
Which MCP servers can execute transactions?
Servers with signing or broadcast tools. See MCP and payments for payment-focused servers.
Related Reading
Sources and Methodology
Tool names, flags and limits are quoted from each project's documentation, fetched on November 12, 2026. No wallet was connected or funded for this page.

