Integration Documentation

Connect InferenceBrake to your AI agents in minutes. The Free plan includes 5,000 checks/month; paid plans scale to 100k and 500k per month.

Runnable examples: InferenceBrake/inferencebrake-examples.

Before You Start

Sign up at inferencebrake.dev
Get your API key from the Dashboard

Python SDK

Install the inferencebrake package and start monitoring your agents.

Install

pip install inferencebrake

Basic Usage

Python

from inferencebrake import InferenceBrake

guard = InferenceBrake(api_key="ib_your_key")

for step in agent.run():
    status = guard.check(
        reasoning=step.reasoning,
        session_id="agent-session-1",
        action=step.tool,       # optional, enables action repetition
        model="gpt-4o-mini",    # optional, recorded for attribution
        prompt="weather task",  # optional, recorded for attribution
    )

    if status.should_stop:
        print(f"Loop detected! {status.detector_triggered}")
        break

Batch Check

Python

statuses = guard.check_batch(
    reasoning_list=["step1", "step2", "step3"],
    session_id="agent-session-1"
)

Session History

Python

history = guard.get_session_history(
    session_id="agent-session-1",
    limit=50
)

Configuration

Python

guard = InferenceBrake(
    api_key="ib_your_key",
    timeout=10,
    auto_stop=False
)

JavaScript / Node.js SDK

Built-in resilience: retry logic, circuit breaker, and offline queue.

Install

npm install inferencebrake

Basic Usage

JavaScript

const { InferenceBrake } = require('inferencebrake');

const guard = new InferenceBrake({ apiKey: 'ib_your_key' });

const status = await guard.check(
    'reasoning text',
    'session-1'
);

if (status.shouldStop) {
    console.log('Loop detected!', status.message);
}

Resilience Configuration

JavaScript

const guard = new InferenceBrake({
    apiKey: 'ib_your_key',
    timeout: 10000,
    maxRetries: 3,
    retryDelay: 1000,
    retryBackoff: 2,
    circuitBreakerThreshold: 5,
    circuitBreakerTimeout: 30000,
});

Monitor Helper

JavaScript

const { inferencebrakeMonitor } = require('inferencebrake');

const monitor = inferencebrakeMonitor({ apiKey: 'ib_your_key' });

const status = await monitor.check(reasoning);

monitor.reset('new-session-id');

REST API

Use directly from any language or framework via HTTP.

Health Check

Shell

curl https://ocnjiyiqeifllbyqohks.supabase.co/functions/v1/health

Check Reasoning

Shell

curl -X POST https://ocnjiyiqeifllbyqohks.supabase.co/functions/v1/check \
  -H "Authorization: Bearer ib_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "session_id": "agent-123",
    "reasoning": "I should check the weather in NYC"
  }'

Response

JSON

{
  "action": "PROCEED",
  "loop_detected": false,
  "similarity": 0.42,
  "confidence": 0.18,
  "detectors": {
    "semantic": false,
    "token_repeat": false,
    "action": false,
    "ngram": false,
    "editdist": false,
    "compression": false
  },
  "estimated_cost_saved": 0,
  "status": "safe",
  "message": "Reasoning sound",
  "usage": {
    "today": 42,
    "month": 42,
    "limit": 5000,
    "remaining": 4958
  }
}

API Reference

CheckStatus Fields

Field	Type	Description
`action`	string	`"KILL"` if loop detected, `"PROCEED"` otherwise
`loop_detected`	boolean	Whether a reasoning loop was detected
`similarity`	float	Max cosine similarity against recent steps (0.0 - 1.0)
`action_repeat_count`	int	Number of consecutive identical actions
`ngram_overlap`	float	N-gram overlap ratio with recent steps
`confidence`	float	Weighted voting confidence (0.0 - 1.0)
`status`	string	`"safe"`, `"warning"`, or `"danger"`
`message`	string	Human-readable status message
`test_mode`	boolean	Whether request used test mode API key

Detector Fields

Detector	Method	Best For
`semantic`	Embedding cosine similarity	Paraphrased repetition
`token_repeat`	Exact repeated token spans	Verbatim loops (Antidoom / OpenRouter failure mode)
`action`	Tool call patterns	Repeated tool invocations
`ngram`	Text overlap	Phrase-level repetition
`editdist`	Normalized Levenshtein	Near-identical mirror loops
`compression`	Normalized Compression Distance	Structural / information theory

Rate Limits

Free plan: 5,000 checks per month per account. Paid plans raise the limit. Counters reset on the 1st.

Rate limit headers are returned with every response:

X-RateLimit-Limit - Monthly limit
X-RateLimit-Remaining - Checks remaining this month
X-RateLimit-Period - month
X-RateLimit-Reset - Unix timestamp when the quota resets

Error Codes

Status	Error	Cause
401	`Invalid API Key`	Missing or invalid `Authorization` header
402	`Subscription past due`	Payment required
429	`Rate limit exceeded`	Daily check limit reached
500	`Internal error`	Server-side failure, retry with backoff

Framework Integrations

LangChain (Python)

Python

from inferencebrake import InferenceBrakeCallbackHandler

handler = InferenceBrakeCallbackHandler(api_key="ib_your_key")

agent = AgentExecutor(agent=agent, tools=tools, callbacks=[handler])

CrewAI (Python)

Python

from inferencebrake import create_crewai_callback

callback = create_crewai_callback(api_key="ib_your_key")

agent.callbacks = [callback]

Python Decorator

Python

from inferencebrake import guard_agent_loop

@guard_agent_loop(
    api_key="ib_your_key",
    session_id="agent-1",
    action=lambda r: r["tool"],
)
def call_model(prompt):
    return client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
    )

JavaScript Monitor

JavaScript

const { InferenceBrakeCallbackHandler } = require('inferencebrake');

const handler = new InferenceBrakeCallbackHandler({
    apiKey: 'ib_your_key',
    sessionId: 'my-agent'
});

const result = await chain.invoke(input, { callbacks: [handler] });

Escalation

Instead of stopping on the first loop, give the agent a chance to recover on a stronger model, then stop if it still loops. The SDK does not pick models; you provide the hook.

Python

from inferencebrake import guard_agent_loop

@guard_agent_loop(
    api_key="ib_your_key",
    session_id="agent-1",
    escalate=lambda status, attempt: switch_model("claude-opus"),
    max_escalations=2,
)
def call_model(prompt):
    ...  # on a loop the agent retries on the stronger model,
         # then stops if it still loops

Order of precedence on detection:

escalate(status, attempt) while under max_escalations
on_loop(status)
raise LoopDetectedError when auto_stop is set

Use steering_message(status) for a ready-to-inject nudge, and LoopPolicy to apply the same behavior outside a decorator.

Analytics

Attribute loops to the model, tool, and prompt that produced them, not just a total count.

cURL

curl -H "Authorization: Bearer ib_your_key" \
  "https://ocnjiyiqeifllbyqohks.supabase.co/functions/v1/analytics-summary?days=30"

Response

{
  "total_checks": 78,
  "loops_blocked": 9,
  "estimated_usd_saved": 0.0146,
  "by_model": [
    { "model": "gpt-4o-mini", "checks": 30, "loops": 4, "saved": 0.006 }
  ],
  "by_action": [
    { "action": "get_balance", "checks": 12, "loops": 3, "saved": 0.004 }
  ],
  "by_detector": [
    { "detector": "semantic", "loops": 6 },
    { "detector": "token_repeat", "loops": 3 }
  ],
  "recent_loops": [
    {
      "session_id": "agent-1",
      "model": "gpt-4o-mini",
      "action": "get_balance",
      "confidence": 0.85,
      "saved": 0.0019,
      "created_at": "2026-09-20T03:24:44Z"
    }
  ]
}

Attribution comes from the model, action, and prompt you pass to check(). Totals reflect retained metrics, so the window is bounded by your plan's retention.