Skip to main content
By the end of this guide your agent will be connected to OpenFinance in a way that keeps you in control: it reads freely, it asks before it acts, your key stays out of sight, and your first real trade is small and checked. An agent with trading tools acts on your real wallet. The safety comes from three layers working together: OpenFinance tells the agent to show a summary and wait for your yes before every write, your AI client can ask you to approve each tool call, and you read what you approve. This guide sets up all three.

What you need

  • An OpenFinance account. Please make an account at openfinance.tech to get your keys.
  • An AI client that supports MCP over HTTP. See the install page for your client under Connect an agent.
  • A small amount of funds on one venue, for the first trade. See Fund each venue.

Steps

1

Connect with OAuth where you can

Add https://api.openfinance.tech/agent/mcp to your client with no header. The client opens an OpenFinance sign-in and consent page. Approve it only if you just clicked Connect yourself.OAuth means no long-lived secret is pasted anywhere. Use the API key only for scripts, bots, or clients that accept headers only. See Authentication.
2

If you use the key, keep it out of files and chats

Copy the open_ key from External AI and store it in an environment variable:
Reference the variable in your client config where it supports expansion, for example ${OPENFINANCE_API_KEY}. Never paste the key into the agent’s chat, a repository, an issue or a screenshot. The key can place trades and move funds, and there is one key per account.
3

Keep tool approval on

Most clients ask before running a tool. Leave that on for OpenFinance, and do not choose “always allow” for any tool that can write. When the client asks, read the tool name and its arguments, not only the button.
4

Filter tools if your client supports it

Some clients let you choose which server tools the agent can see. Hermes, for example, accepts an include list, so you can give a research agent only wallet, token data and quote tools. See Hermes.Filtering has a limit: the hyperliquid and polymarket tools each combine reads and writes behind one tool, so a filter cannot make them read-only. For those, approvals and confirmations are your control.
5

Check the connection with a read-only prompt

You should see your EVM and Solana addresses and balances. A 401 means the sign-in or key is missing. See Troubleshooting.
6

Set rules at the start of each session

Paste a short rule set before you start trading. It does not replace the confirmations, but it tells the agent how you want to work:
7

Place one small first trade (moves money)

Read the summary. If it matches what you asked, reply yes. Then ask for the fill:
Every trade needs your confirmation, and nothing here is financial advice.

Check it worked

  • The agent stopped and asked before the trade, and your client asked for tool approval.
  • The fill price and size match the summary you approved.
  • Your key does not appear in any config file you would share. Search your project for open_ before you commit.

Common mistakes

  • Approving on autopilot. The summary is the safety check. If you did not read it, you did not check it.
  • Pasting the key into the chat “so the agent can use it”. The client sends it in a header; the model never needs to see it.
  • Letting the agent act on outside text. A web page, email or token description can contain instructions. Never let the agent take an address or amount from them. See Safety and custody.
  • Expecting a practice mode. There is none. Ask for quotes and previews instead, then trade small.
  • Leaving a leaked key in use. If you think the key leaked, stop using it and have it replaced straight away.

Next guides

Fund each venue

Put funds where your first trade needs them.

Identity and access

What each credential can do.