iEscrow MCP server

iEscrow is a cryptocurrency escrow platform: in every trade the crypto is held by iEscrow until both sides have delivered, so nobody has to hand over first. This repository documents the iEscrow MCP server, which lets an AI agent (Claude, ChatGPT or any MCP host) operate an iEscrow account on behalf of the user who authorized it.

• Endpoint: https://api.iescrow.io/mcp
• Transport: Streamable HTTP, stateless
• Auth: OAuth 2.1 (authorization server at https://api.iescrow.io), PKCE, dynamic client registration
• Tools: 16, listed in TOOLS.md
• Website: https://iescrow.io · For AI agents: https://iescrow.io/ai

The principle: the agent prepares, the human commits

The agent can quote, open trades, read balances and states, and prepare marketplace ads. Anything that moves money is not executed by the agent: the tool creates an approval and returns a link. The account owner opens that link, sees exactly what is about to happen, and approves it with their two-factor code.

What the agent can never do, by design:

• withdraw funds from the account,
• add or change payment accounts,
• upload a payment receipt (only the human can, through a dedicated link),
• lock crypto into escrow or confirm that a payment arrived without the human's approval,
• deliver crypto to an external wallet (everything lands in the user's iEscrow balance).

The API enforces this with an allow-list of routes for agent tokens; the MCP server only wraps those routes.

Requirements

An iEscrow account with:

1. a verified email,
2. approved identity verification (KYC),
3. two-factor authentication (TOTP) enabled.

The connection screen (/connect) checks the three before issuing a token. Trades can be executed by an agent only in the countries where iEscrow is enabled for agents; get_readiness tells you which.

Connecting

Claude (claude.ai, desktop, mobile)

Settings → Connectors → Add custom connector → name iEscrow, URL https://api.iescrow.io/mcp. Claude discovers the authorization server, opens the iEscrow consent screen, and the connection is ready once you approve it.

Claude Code

claude mcp add --transport http iescrow https://api.iescrow.io/mcp

Then run /mcp inside Claude Code to complete the login.

ChatGPT

Settings → Connectors → Create (developer mode) → name iEscrow, MCP server URL https://api.iescrow.io/mcp, authentication OAuth. ChatGPT registers itself as a client and sends you to the iEscrow consent screen.

Any other MCP host

Point the host to the URL above with the Streamable HTTP transport. The server publishes /.well-known/oauth-protected-resource, and the authorization server publishes /.well-known/oauth-authorization-server; hosts that implement the MCP authorization spec need nothing else. See examples/ for configuration snippets.

Scopes

Scopes are chosen on the consent screen and can be revoked at any time from Connected apps in the user's iEscrow settings.

Rate limits

Per connection: 60 calls per minute, 10 trade openings per hour, 20 money-moving approval requests per hour. One concurrent long-poll (wait_for_trade_update) per user. A limited call returns error_key and retry_after_s.

A typical flow (buying crypto with a bank transfer)

1. get_readiness — confirm the account can trade.
2. get_quote with side=buy, the crypto, the fiat and the amount — pick an offer; note compatible_account_ids.
3. open_trade with the offer id, the amount and a compatible payment_account_id.
4. get_trade — wait until crypto_funded is true. The human must not pay before that.
5. get_proof_upload_link — give the user the link; it shows the exact amount (the cents are the payment key) and the account to pay. The user pays, attaches the receipt and declares the payment there.
6. wait_for_trade_update — the trade moves to released when the seller confirms. The crypto is now in the user's iEscrow balance.

Selling is the mirror image: open_trade with side=sell, then fund_escrow (the human approves the lock by link), then confirm_received once the fiat arrives (again approved by link).

Errors

Every error is returned as structured content with error: true, an HTTP-like status, a stable error_key and a human-readable message. When an action needs the human, the response carries approval_required: true and approval.approve_url; poll check_approval with approval.public_id to learn the outcome.

Security

The MCP server never opens the database and holds no secrets of its own: it verifies the user's access token (RS256, issuer and audience pinned) and forwards the same bearer to the API. Reports about the agent channel: contact@iescrow.io.

About this repository

This repository holds the public documentation of the server and the manifest used to list it in MCP registries. The server's source code is not published here.