Skip to main content
Today the Base MCP exposes a single execution mode for write tools: approval mode (the user manually approves each transaction via a returned URL). Other modes may be added later. Treat the tool descriptions exposed by the MCP as the source of truth — if a future write tool returns a different shape or skips the approval step, follow what the MCP describes, not this file.

​
Approval Mode

In approval mode, every write call (send, swap, sign, batched calls, and any plugin-prepared transaction routed through Base MCP) returns an approval URL plus a request ID. The user opens the URL, approves the action in Base Account, and then the agent polls the request ID for completion.

​
Flow

  1. Call the write tool. The response includes:
    • an approval URL (the field name is on the MCP response — typically approvalUrl)
    • a request ID (typically requestId)
  2. Show the user the link. Present it as “Approve Transaction” (or similar neutral language). Refer to the approval destination as Base Account, not as the raw URL hostname or an implementation-specific provider. Just give the user the link to click.
  3. In CLI harnesses, also open the link automatically. When you’re running in an environment with a Bash/shell tool (Claude Code, Codex, Cursor terminal, etc.), don’t just print the URL — also open it in the user’s default browser so they don’t have to click. Always print the link too as a fallback, then run the platform-appropriate open command:
    • macOS: open "<url>"
    • Linux: xdg-open "<url>" (fall back to wslview under WSL)
    • Windows: start "" "<url>" (or via PowerShell: Start-Process "<url>")
    Skip this step on chat-only surfaces (ChatGPT, Claude.ai) — they don’t have a shell, so just show the link.
  4. Wait for the user to confirm they approved. Don’t poll in a tight loop while they’re still acting.
  5. Call the status-poll tool (typically get_request_status) with the request ID once.
  6. Only report success when the status tool confirms completion.

​
Common mistakes

  • Reporting success before the status tool confirms it — the user may not have approved yet.
  • Skipping the approval link — the transaction cannot complete without user action.
  • Naming the wallet/approval provider, or surfacing the raw hostname as the link text — say “Approve Transaction”.
  • Polling the status tool in a tight loop instead of once after the user confirms.
  • Forgetting to also auto-open the link in CLI harnesses where a shell is available — printing alone makes the user copy-paste unnecessarily.
  • Trying to auto-open in chat-only harnesses where no shell exists — that just produces an error.