Mercury

No. Mercury states its hosted MCP is limited to read-only actions and cannot initiate transactions or modify account data. All 31 documented tools are get or list operations. Mercury's payment endpoints, createTransaction and requestSendMoney, exist in the REST API but are absent from the MCP tool list, and the OAuth token is issued with a read scope.

No, and nothing in Mercury's public API can. The MCP tool listSendMoneyApprovalRequests only reads the queue of pending payments. Mercury publishes no approve or reject endpoint anywhere in its API reference, so approval happens exclusively in the Mercury web interface by a human with send-money permission. The connector can see a pending payment but not act on it.

The two lists are entirely disjoint and one of them is stale. Anthropic's snapshot from 2026-01-12 shows five snake_case names such as load_account, none of which appears anywhere in Mercury's documentation. Mercury's own tool page, updated 2026-07-27, lists 31 camelCase names. Treat the vendor list as current and the directory as an old naming scheme.

Two, and both are narrow. Mercury's RFC 9728 resource descriptor declared scopes_supported of read and offline_access on 2026-08-22. The read scope is the entire data grant, and offline_access only lets a client refresh its token without sending you back to the browser. There is no write scope to grant, which is why the boundary holds.

Full bank routing and account numbers for your payment recipients, plus balances, complete transaction history with counterparties, statements and your EIN. Mercury's own security guidance names recipient routing numbers and account information explicitly. Card tools return only the last four digits. Read-only does not mean low risk here — the read surface is the whole risk.

No. Mercury states there is nothing to set up on its side and that your AI tool registers itself the first time it connects, using OAuth 2.0 Dynamic Client Registration. Adding the server grants no access on its own; Mercury issues a token only after you sign in and select Allow at its own consent screen.

The session expired and needs a fresh sign-in. Mercury states a Claude connector session lasts about three days on the same chat thread. Command line tools last longer because Mercury returns a refresh token to any client that requests the offline_access scope. Run claude mcp login mercury to sign in again.

Mercury labels it Beta. Its documentation carries a warning that the MCP is in Beta while the team understands the limitations of chat models, and asks you to double check any response from the model against your Mercury account before making important decisions. Treat figures the agent reports as a draft, not as a reconciled statement.