Local MCP server

Your AI agent is blind after you click Sign.

WalletBridge is an MCP server that gives agents live onchain context — simulate, decode, read — without your keys leaving your machine. No daemon. No extension. Just an MCP server your agent already speaks.

MCP stdio — JSON-RPC session
$ npx walletbridge
  // stdio server running. agent connects via MCP.
 
// simulate a swap before signing
> simulate_transaction
  fork: ethereum @ block 21,403,112
  from: 0x7a3d...0000  |  to: UniswapV3: Router
  value: 0.5 ETH → USDC
  ✓ success — output: 1,847.32 USDC
 
// check the agent's wallet balance first
> get_account_state
  balance: 4.2819 ETH  |  nonce: 14  |  type: EOA
 
$ _
No extension. No daemon. Just an MCP server your agent already speaks.

How it works

RPC
any chain · real state · no sandbox
→
Anvil Fork
fork at tx block · simulate without signing
→
viem
decode · format · normalize ABI
→
MCP Server
stdio · no network · agent-native
→
Agent
Claude Code · Cursor · any MCP client

Four tools. That's the whole API.

simulate_transaction

Fork-based execution. Revert is decoded if it fails — the agent sees exactly what went wrong and why, no guessing.

resolve_abi

Fetches from Etherscan on demand, falls back to provided ABI. Results cached in-memory for the session.

read_contract

Onchain state reads — balances, allowances, storage slots. Works across any EVM chain the agent has an RPC URL for.

get_account_state

Balance, nonce, code-type detection (EOA vs deployed contract). Fast — no fork required.

The feedback loop problem

You click Sign. The agent goes silent. The tx fails. You get a raw revert. You have no idea if the agent understood why.

WalletBridge closes the action→result feedback loop. The agent simulates before signing, decodes failures, reads state — all before you touch the transaction.

No cloud. No keys leave your machine. No daemon. Just a local MCP server.

1
tool that changes everything: simulate_transaction on a fork
0
external dependencies — no browser extension, no daemon
4
MCP tools. Nothing else to learn.

Stop building AI agents that work blind onchain.

Close the action→result feedback loop. Your keys never leave your machine.