How it works
A checkpoint between your agent and your money.
Railguard sits between agent intent and your signer. It decides before anything is signed, watches what happens after, and leaves a receipt. It never takes custody.
Your agent→Railguard: check · reserve · watch · prove→Your wallet or bank API→Chain or bank
The seven steps of every payment
- 1Intent
- The agent says what it wants to pay: who, how much, which coin, for what task.
- 2Authorize
- Your rules decide: allowed recipients, budgets, whether a person must approve.
- 3Reserve
- The money is set aside first, so two agents cannot spend the same budget.
- 4Execute
- Your own wallet or provider signs and sends. Railguard never holds keys.
- 5Observe
- Railguard watches the chain or bank API. A lost answer is UNKNOWN, not failed.
- 6Reconcile
- What the chain shows is matched to the intent. Any difference is flagged.
- 7Evidence
- A hash-chained receipt records every step, so anyone can re-check it.
Words used on this site
- Intent
- What the agent wants to pay: who, how much, which coin.
- Policy
- Your rules: budgets, allowed recipients, approvals.
- Reserve
- Money set aside before sending, so it cannot be spent twice.
- UNKNOWN
- Sent, but no answer yet. Railguard checks instead of guessing.
- Confirmed
- The chain shows the payment matched the intent (settlement verified).
- Receipt
- A tamper-evident record of every step that you can re-check.
- Mandate
- An on-chain spending limit for one agent (used on Metal).
- x402
- A way for agents to pay for an API call over HTTP, per request.
- Soroban
- Stellar's smart contract platform, used for the on-chain policy.
- Testnet
- A practice blockchain: fake money, real code paths.
Run it yourself
Needs Bun. Lines starting with $ go in a terminal.
1. Install and attack
Clone the repo, then run the six failure tests locally.
$ git clone https://github.com/prasanthkuna/railguard-gateway.git $ cd railguard-gateway && bun install $ bun run railguard attack $ bun run railguard protect
2. Re-check the public proofs
Verifies the committed evidence for each chain.
$ bun run verify-arbitrum-sepolia-pack $ bun run verify-stellar-testnet-pack $ bun run verify-monad-metropolis-pack
3. Give your agent Railguard tools (MCP)
Paste into your Cursor or Claude Desktop MCP config, not a terminal.
{
"mcpServers": {
"railguard": {
"command": "bun",
"args": ["run", "packages/mcp/src/server.ts"],
"env": {
"RAILGUARD_BASE_URL": "http://127.0.0.1:4000",
"RAILGUARD_ACCESS_TOKEN": "<your-token>"
}
}
}
}4. Call it from code
TypeScript SDK over the REST API (/v1/*).
import { RailguardClient, check } from "@railguard/sdk"
const rg = new RailguardClient({ baseUrl, getAuthHeaders })
const result = await check(rg, intent)
if (result.decision === "ALLOW") await rg.execute(result.intent.id)How it is kept honest
- On every push to mainlint, unit tests, lifecycle tests, manifest drift check, builds
- Evidence workflowtestnet evidence packs re-verified in CI
- Status labelsgenerated from one manifest, checked against the docs
- Maturityv0.1 alpha, testnet only, not for mainnet funds yet