The lightweight, developer-first safety layer for AI actions
Queue, approve, and execute API calls safely from your agents. Add human approval, retries, and audit trails in minutes.
If your AI can issue refunds, update customer data, or trigger internal APIs — one bad output hits a real API. One retry becomes a duplicate charge.
SafeFetch sits between your agent and the real world.
Replace any API call your AI makes with SafeFetch.
Every action goes through a gate.
You decide what needs a human approval and when to skip approval for trusted endpoints.
// Your agent calls SafeFetch, not Stripe
send_action({
url: "https://api.stripe.com/v1/refunds",
method: "POST",
body: { charge: "ch_3Qx9R2eZ" },
approve: true,
})
// => { id: "act_Xk9mP2nQ4rT1vW", status: "awaiting_approval" }
// You approve. The action executes.
approve({ id: "act_Xk9mP2nQ4rT1vW" })
// => { status: "completed", response_code: 200 } Your AI doesn't call Stripe directly. It goes through SafeFetch.
Your AI sends actions to SafeFetch instead of calling APIs directly.
Refund a payment, update a CRM record, trigger a deploy. Instead of calling the API directly, it calls SafeFetch.
The action is validated, deduped, and held for approval if configured. Retries and signing are automatic.
Approve via MCP, SDK, CLI, or API. Or skip approval for trusted endpoints.
Signed delivery to the target API. Full response stored. Your agent or system can fetch the result anytime.
Because this happens.
Before
// ❌ BEFORE — your AI calls Stripe directly
await stripe.refunds.create({
charge: "ch_3Qx9R2eZ",
})
// No approval. No retries. No audit trail.
// A hallucinated charge ID hits a real API.
// A retry double-refunds a customer. After
// ✅ AFTER — your AI goes through SafeFetch
import { safeFetch } from "safefetch"
await safeFetch({
url: "https://api.stripe.com/v1/refunds",
method: "POST",
body: { charge: "ch_3Qx9R2eZ" },
approve: true,
})
// Human approves. SafeFetch executes.
// Signed delivery. Full audit trail. Safe retries. fetch() with safeFetch()You don't need to re-architect anything. Import the SDK, swap one function, and every API call your AI makes now has approval, retries, deduplication, and a full audit trail.
fetch(). Just add approve: true.import { safeFetch } from "safefetch"
// Before: await fetch("https://api.stripe.com/v1/refunds", ...)
// After:
await safeFetch({
url: "https://api.stripe.com/v1/refunds",
method: "POST",
body: { charge: "ch_3Qx9R2eZ" },
approve: true, // hold for human approval
dedupe: "refund-ch_3Qx9R2eZ", // prevent duplicates
}) import { safeFetch } from "safefetch"
// Define tools that go through SafeFetch.
// The agent never gets direct API access.
const refundTool = {
name: "refund_payment",
description: "Refund a Stripe charge",
execute: async (args) => {
return safeFetch({
url: "https://api.stripe.com/v1/refunds",
method: "POST",
body: args,
approve: true,
})
},
}
// Your agent only knows about refund_payment.
// It cannot call Stripe directly. The strongest safety pattern: your agent doesn't have credentials to call production APIs. Only SafeFetch does. Define tools that route through SafeFetch, and your AI literally cannot bypass the safety layer.
From AI support agents to internal copilots. SafeFetch handles actions so your AI doesn't go rogue.
Your support AI handles refunds, credits, and account changes. Every sensitive action held for approval — so a hallucinated charge ID doesn't trigger an unintended refund, and a retry doesn't double-charge a customer.
→ Require approval before execution. Auto-approve safe reads.
Modify CRM, run admin actions.
→ Audit everything your AI does
Sync systems, call flaky APIs.
→ Retries + durability built in
A complete safety layer. Simple REST API, SDK, MCP tools, and CLI.
Pause actions for sign-off. Approve via API, MCP, or email magic link. Auto-approve safe reads.
Dedupe keys and idempotency headers. No double refunds from agent retries.
Exponential backoff from 20s to 1h. Configurable max attempts per action.
Every request signed with HMAC SHA-256. Your target API can verify it came from SafeFetch.
Every action logged — who approved it, when it ran, the full response, duration, and status. Queryable via API.
Get notified when actions complete or fail. POST to any callback URL with the full result.
Works with MCP, SDK, CLI, and REST API. Connects to ChatGPT, Claude, Cursor, and any agent framework.
Your agent can queue actions, request approval, and check results — all through MCP tools. No custom code needed.
Per job, not per message. Safety features on every plan.
Experiment safely
Start freeProduction protection
Get startedEnterprise-grade control
Get startedAll plans include approval gates, safe retries, deduplication, signed delivery, stored results, audit history, SDK, MCP, and CLI.
No credit card required. Get your API key in 30 seconds.
Your AI should never call production APIs directly.