Same execution guard — one lens: any irreversible API call

Execution Integrity for Irreversible APIs

Retries are ambiguous. Duplicate side effects aren't. SafeAgent durably claims a logical action and keeps uncertain outcomes blocked until they are reconciled. Payments, webhooks, and agent tool calls still require provider evidence for end-to-end outcome certainty.

Install from PyPI View on GitHub See it happen ↓

pip install safeagent-exec-guard

Python 3.10+ Apache-2.0

01 — Where it applies

Any API where "it happened" and "the response arrived" aren't the same fact.

The gap between an action firing and your code confirming it fired is where duplicates come from. These are the categories where that gap has a direct dollar or trust cost.

Payments

Stripe, Coinbase, Binance, and broker APIs (Alpaca, Interactive Brokers) all execute irreversible transfers of money. A retried charge, trade, or transfer isn't a UI glitch — it's a second real-world transaction.

Stripe

Coinbase

Binance

Alpaca

Interactive Brokers

Webhooks

Stripe, GitHub, and Twilio explicitly document at-least-once delivery — they will redeliver an event your endpoint doesn't cleanly acknowledge. Slack's Events API behaves the same way, retrying on anything but a fast 200. None of them guarantee exactly-once. That's the receiving side's job.

Stripe

GitHub

Twilio

Slack

Agent tool calls

OpenAI tool calls, Anthropic/Claude tool use, and any agent framework with side-effecting actions can be retried by the orchestration layer without knowing whether the underlying action already executed.

OpenAI tools

Claude / MCP

LangChain

CrewAI

n8n

02 — See it happen

One button. Two outcomes.

This is the actual failure mode — mash the button and watch one side fire a duplicate order on every click while the other side blocks every repeat after the first. Full interactive demo, live below.

Open full demo in a new tab →

03 — Production proof

Live duplicate blocking — May 21, 2026.

Not a simulation. Six confirmed SKIP events from a live trading session on the full stack: DashClaw, SafeAgent, Mycelium Trails, Base/Arbitrum, broker Alpaca.

Six duplicates blocked in one live session

09:42 ET — duplicate buy TQQQ qty=6 blocked, $452

09:47 ET — duplicate add TQQQ qty=6 blocked, $452

09:49 ET — duplicate sell TQQQ qty=12 on flip blocked, $902

10:00 ET — duplicate entry TQQQ qty=6 blocked, $454

10:14 ET — duplicate sell TQQQ qty=18 on V20 flip blocked, $1,350

11:06 ET — duplicate SQQQ add during scale-in blocked, $43

Sum of the six blocked events above: $3,653 in duplicate exposure prevented in a single session.

Full session data →

Add the guard before the next retry fires.

MIT-licensed claim-before-execute guard. About 10 lines to add to any agent, workflow, or webhook handler.

View on GitHub

pip install safeagent-exec-guard