For developers
An MCP server that gives your agent money. And keeps it inside the rules.
Lotvs runs a remote Model Context Protocol server. Any client that speaks MCP over Streamable HTTP can connect, sign in with OAuth or use an access key, and act on an account of its own, with the owner's limits checked on the server.
Free to sign up. Facts on this page checked against the product on October 16, 2026.
Connect
One address, any MCP client
https://www.lotvs.ai/api/mcpclaude mcp add --transport http lotvs https://www.lotvs.ai/api/mcp{
"mcpServers": {
"lotvs": {
"url": "https://www.lotvs.ai/api/mcp"
}
}
}POST https://www.lotvs.ai/api/mcp
Authorization: Bearer <access key>An unauthenticated request gets 401 with a WWW-Authenticate header pointing to the OAuth metadata (OAuth 2.1 with PKCE and dynamic client registration).
Tools
What an agent can call
Every tool acts on the calling agent's own account. Tools that move money are checked against the owner's rules before anything moves.
get_agent_account- Balance, permissions, limits left and whether it is paused. Read only.
list_payments- Recent payments and trades, newest first. Read only.
list_assets- What it may buy and sell. Read only.
pay_usdc- Send USDC to the owner or to a person the owner allowed. Moves money.
pay_x402- Fetch an https URL that charges through x402 and pay its price, up to max_usdc. Moves money.
check_request- How a payment that waited for the owner ended.
buy_asset, sell_asset- Trade an asset from the catalog, within the limit per trade. Moves money.
earn_deposit, earn_withdraw- Lend spare USDC in Earn and take it back. Moves money.
create_vault, move_to_vault, move_from_vault- Set money aside inside the agent's account, when the owner allows it.
Full parameters and annotations are in the tools reference.
Design
Built for agents that can be wrong
Rules before anything moves
Limits, allowed recipients and approvals are enforced before anything moves, never by the model.
Errors an agent can use
A refusal is one sentence with the rule and the figures, returned as a tool error. Nothing to work around.
Approvals without double payment
A payment that needs the owner's yes returns a request id; the agent checks back instead of paying again.
No money in the model
The agent never holds the money or an address. Only what passes the owner's rules goes through.
Questions
Straight answers
Which transport does the server use?
Streamable HTTP at one address. Clients that support OAuth sign in the same way Claude does; others use an access key as a bearer token.
Can one agent act on another agent's account?
No. Each connection acts on its own account only.
What does a refusal look like?
A tool result with isError true and one sentence naming the rule and the figures, for example a payment over the limit per payment. The owner sees the attempt too.
Is there a sandbox?
Not for agent accounts today. Start with small limits and a small balance.
In the documentation
Related
- Connect ClaudeAdd Lotvs to Claude as a connector in one click. Claude gets an account of its own to pay, invest and earn within limits you set, on the web, desktop and phone.
- Connect ChatGPTAdd Lotvs to ChatGPT as a custom MCP server. ChatGPT gets an account of its own to pay, invest and earn within limits you set.
- x402 paymentsLet your AI agent pay for APIs and sites that answer HTTP 402 Payment Required, in USDC, within a price it sets and your limits. And get paid by agents with @lotvs/pay.
- Spending limits for AI agentsSet how much your AI agent may spend per payment, per day and per trade, who it may pay, and when it must ask you first. Lotvs checks every rule before money moves.
Important
Your agents' money is kept apart from your own account, with each agent's share in Lotvs's records. The AI never holds the money or an address: it can only ask, and only what passes your rules goes through.
Lotvs is a financial technology company, not a bank. Balances are held as digital dollars, not bank deposits, and are not covered by deposit insurance. See the Terms and Privacy Policy.