spendctl sandbox: a hosted MCP server where agents pay for APIs under a budget
nomunomu0504 · reddit · 2026-10-27
The author released a hosted sandbox for spendctl, a spend-policy sidecar for agents that pay per API request via x402/MPP, exposed as a remote MCP server with no login and fully mocked internals.
Key features:
- Per-agent budgets, host allowlists, recipient pinning, price-jump checks, human approval above a threshold, and a receipt ledger
- The agent never holds API keys; a sidecar signs and enforces payments
The sandbox exposes six tools (sandboxcreate, sandboxpay, checkbudget, listreceipts, getpolicy, approve) with a throwaway wallet, mock facilitator, and two fake paid APIs—no outbound network, deleted after 2 hours idle. Works with Claude Code, Cursor, and VS Code via https://mcp.spendctl.dev/mcp.
The real version runs locally with npx spendctl mcp, free for two agents, paid tiers for teams; the sidecar itself is closed source, which is why the sandbox exists—to let you watch exactly what it refuses before installing.
The author also asks whether a convention is emerging for "tool call waiting on human approval": held payments used to return isError: true (reported by Claude as failures) and now return a normal result with held and an approval id.
More from coding & agent
- Claude Opus 4.5 + UE5 re-renders Wuthering Waves environments — Eric520CC · 2026-10-27
- First real agent disaster in 8 months: OpenBot overrides user's declined irreversible setting — Teknium · 2026-10-27
- Renoise MCP Lets Codex and Claude Cut Promo Videos With 30 Built-in AI Models — vista8 · 2026-10-27
- AEGIS open-sources an MCP gateway requiring per-call approval of exact tool arguments — Leather_Tour_794 · 2026-10-27
- Dev says stop underestimating agents using Figma as a planning tool — nathanborror · 2026-10-27
- Dev mod makes GitHub Copilot speak its replies via Microsoft's MAI-Voice-2-Flash — PaulShellDev · 2026-10-27