Runtime safety
for AI agents.

Loop detection · Budget enforcement · Side-effect tracking · EU AI Act compliance

One decorator. Works with LangGraph, CrewAI, Google ADK, raw SDKs, and any MCP server.

Quickstart

Also available: pip install fuze-ai

EU AI Act

Enforcement starts August 2, 2026.

Autonomous agents that make decisions, process personal data, or interact with external systems may qualify as high-risk AI. The penalties are not hypothetical.

€35M

or 7% global turnover

Maximum penalty per serious violation of the EU AI Act.

72h

incident report window

Time to report a serious incident to national authorities under Art. 73.

Aug 2026

full enforcement

High-risk system obligations become fully enforceable. 4 months away.

Every major agent framework — LangGraph, CrewAI, Google ADK — saves state but leaves failure detection, audit logging, and human oversight entirely to you. Fuze fills that gap in one decorator.

See compliance coverage

The Solution

One decorator. Full protection.

Wrap any tool function. Fuze handles loops, budgets, side-effects, and audit trails automatically.

import { guard } from 'fuze-ai'

const search = guard(async (query: string) => {
  return await vectorDb.search(query)
})

const sendInvoice = guard(
  async (id: string, amount: number) => {
    return await stripe.createInvoice(id, amount)
  },
  { sideEffect: true, compensate: cancelInvoice }
)

Features

Everything your framework doesn't do.

Loop Detection

5-layer detection: iteration caps, hash dedup, cost velocity, progress checks, semantic similarity. Catches ping-pong patterns your framework misses.

Budget Enforcement

Hard $/token/time ceilings per step and per run. Pre-execution cost estimation. The run dies before you overspend.

Side-Effect Tracking

Knows which calls changed the real world. Idempotency keys. Compensation functions on rollback. No more double payments.

Smart Recovery

Not just "kill" — retry with modified prompt, rollback to checkpoint, fork to alternate path, or escalate to human.

Audit Trails

Full decision trace: what the LLM saw, decided, called, and what happened. Replayable. EU AI Act Art. 12 compliant.

Framework Agnostic

Works with LangGraph, CrewAI, Google ADK, raw OpenAI/Anthropic SDK, or any MCP server. One decorator, no migration.

How It Works

Three steps to safe agents.

01

Wrap your tools

Add the guard decorator to any tool function. One line of code. No changes to your agent logic.

const search = guard(vectorDb.search)
02

Run your agent

Fuze intercepts every tool call. Tracks costs, detects loops, records side-effects, maintains audit trail.

agent.run(task) // Fuze watching
03

Fuze handles the rest

Budget exceeded? Run halts. Loop detected? Prompt modified or checkpoint restored. Side-effect? Compensated on rollback.

// No double payments. Ever.

Why Fuze

What your framework leaves out.

LangGraph, CrewAI, and Google ADK handle orchestration. They leave safety entirely to you. Fuze fills the gap.

Capability
Fuze
Framework

Loop detection

5-layer: caps, hash dedup, cost velocity, progress, semantic

Budget enforcement

Hard ceilings per step and per run. Pre-execution estimation.

Side-effect tracking

Idempotency keys, compensation functions on rollback.

Audit trail

Full decision trace. Replayable. EU AI Act compliant.

Human escalation

Automatic on budget breach, loop, or unrecoverable state.

State persistence

Fuze adds checkpoints on top of framework state.

Tool orchestration

Fuze wraps framework tools, not replaces them.

MCP Support

Works with any MCP server.

Drop-in proxy for Model Context Protocol servers. Same tool interface, full Fuze protection layer underneath.

Before
// Before: raw MCP server
const server = new MCPServer({
  tools: [searchTool, writeTool, emailTool]
})

// No loop detection
// No budget limits
// No side-effect tracking
// No audit trail
After
// After: Fuze MCP proxy
import { createFuzeMCPProxy } from 'fuze-ai'

const server = createFuzeMCPProxy({
  tools: [searchTool, writeTool, emailTool],
  budget: { maxCost: 0.50, maxTokens: 50_000 },
  sideEffects: { write: true, email: true },
})

// Full protection. Same interface.

Get Started

Ship safer agents today.

One decorator. No migration. Works with your existing stack. EU AI Act compliant out of the box.

Read the docs View on GitHub

MIT License · TypeScript + Python · Zero lock-in