Raposa vs OpenAI Agents SDK
The OpenAI Agents SDK can require approval for a tool call and pause the run. Raposa is a framework-agnostic approval layer that adds the human channels and a verifiable audit trail. Here is the honest difference.
The short version
The OpenAI Agents SDK supports human-in-the-loop by marking a tool as needing approval: the run surfaces an interruption, and your code approves or rejects before the tool executes. A similar pattern exists when exposing tools over MCP (for example with FastMCP), where a call can require confirmation. In both cases the mechanism is in-process — your application decides how the human is asked and what, if anything, is recorded.
Raposa is not tied to a single SDK. It is a dedicated approval layer any agent calls over HTTP, so it works from the OpenAI Agents SDK, another framework, or a plain script. It supplies the pieces the SDK leaves to you: signed approval channels people use from a phone, named approver groups with N-of-M sign-off, reminders and escalation, and a hash-chained, exportable audit trail hosted in the EU.
Side by side
OpenAI Agents SDK and FastMCP details reflect their public documentation as of September 2026. Where their docs leave something to the developer, we write "you build it" rather than imply it cannot be done.
| Dimension | Raposa Aval | OpenAI Agents SDK (+ FastMCP approval) |
|---|---|---|
| Category | Framework-agnostic human approval layer (HTTP API) | Agent SDK with in-code tool-approval / interruption |
| HITL approval scope | One API call gates any action from any framework | Tool calls flagged as needing approval inside the run |
| Who asks the human | Raposa routes to a named approver | Your application code / UI |
| Approval channels | Signed email links, Telegram card, Slack message | You build the channel |
| Named groups & N-of-M sign-off | Built in, every signer recorded | You build it |
| Audit trail verifiability | Hash-chained, signed, exportable, tamper-evident | Your own logs; not a cryptographic audit artifact |
| Silence handling | Timeout is never an approval; reminder, escalation, audited expiry | Whatever your code does while waiting |
| Deploy / payment / comms gating | Same one call for any high-stakes step | Any tool you flag for approval |
| Hosting & data location | EU only (Germany), DPA available | Your infrastructure; model calls via OpenAI |
| Framework coupling | None — call from any stack | OpenAI Agents SDK |
When to choose which
Choose the SDK's built-in approval if you are inside the OpenAI Agents SDK, the reviewer is you or your own UI, and a code-level pause on a tool call is all you need. It is the most direct way to gate a tool within that runtime.
Choose Raposa — or call it from inside an approval-required tool — when the approvers are non-engineers who need email, Telegram or Slack buttons, when you need named groups and two-signature sign-off, and when you need a verifiable, EU-hosted audit trail for compliance. It also stays in place if you change or mix frameworks.
Give your OpenAI agent real approval channels
Call Raposa from an approval-required tool with one HTTP request. Start free — 100 approvals a month, full audit export, no card.