Open source · MIT

Keep a human in the loop on what your AI agents do

Open Approvals lets an AI agent pause and ask before it does something consequential. A shared approval queue, an agent-side client, and a simple protocol any approver surface can implement.

View on GitHub Read the protocol

Agents are starting to take
real, irreversible actions

As agents move from drafting text to doing things, some actions deserve a human’s sign-off first. Open Approvals gives your agent one line to pause on and a place for a person to say yes or no.

Send a messageSpend moneyRun a shell commandDelete dataDeploy to productionEmail a customer

What’s in the box

An approval layer, not a framework

Agent-side client

One call to create an approval request, then block until a human answers. Get back a clear decision: accept, decline, cancel, or timeout.

Shared approval queue

Requests and decisions live in one atomically-written store. Any process on the machine can raise a request or record a decision.

Documented protocol

A small, versioned protocol covers the store format and HTTP endpoints, so any approver surface or agent runtime can speak it.

Approver surfaces

Headless React hooks in open-approvals-react let a dashboard, desktop app, or phone list pending requests and tap Approve or Deny.

Why it’s built this way

Small, unopinionated, and easy to adopt

Open Approvals is a thin layer with zero runtime dependencies. It does one thing — get a human decision on an agent action — and stays out of the way of whatever framework, runtime, or UI you already use.

The agent pauses and asks before it does something consequential

Requests and decisions live in one shared, atomically-written queue

Any surface — dashboard, desktop app, or phone — can approve or deny

Requests from external agent runtimes and MCP elicitations share the queue

The loop

Request, review, resume

Request

The agent asks

Before a sensitive action, the agent creates an approval request and blocks — describing what it wants to do and why.

Review

A human decides

The request appears on any approver surface. A person reviews the details and taps Approve or Deny, wherever they are.

Resume

The decision flows back

The agent unblocks with the decision — accept, decline, cancel, or timeout — and continues, or safely backs off.

Install

One call to keep a human in the loop

agent.py

from open_approvals import create_approval_request, wait_for_approval

# Before a sensitive action, ask a human.
request = create_approval_request(
    requester="billing-agent",
    action="refund",
    description="Refund $240 to customer #1032",
)

decision = wait_for_approval(request["id"], timeout_seconds=300)
if decision["accepted"]:
    issue_refund()

Python (the agent side)

pip install open-approvals

React (the approver side)

npm install open-approvals-react

Read the protocol →open-approvals-react →

Give your agents a checkpoint

Open Approvals is free and open source under the MIT license. Drop it in wherever an agent needs a human’s yes before it acts.