WarderBox

The safety layer between AI agents and the real world.

WarderBox intercepts every tool call your AI agent makes and provides rollback, cost control, smart model routing, cost anomaly detection, rate limiting, circuit breakers, approval workflows, distributed tracing, chaos testing, and compliance reporting — so agents can work autonomously without causing damage.

Zero mandatory dependencies. One pip install. Works with every major agent framework.

Market context: With Guardrails AI acquired by Harvey (Sept 2026) and the AI agent safety market projected to grow from $610M to $6.85B at 41% CAGR, WarderBox fills the gap as a comprehensive, framework-agnostic agent operations platform.

Why WarderBox?

Install

pip install warderbox

With framework integrations:

pip install warderbox[openai]      # OpenAI Agents SDK
pip install warderbox[anthropic]   # Anthropic Claude SDK
pip install warderbox[langchain]   # LangChain
pip install warderbox[crewai]      # CrewAI
pip install warderbox[autogen]     # Microsoft AutoGen
pip install warderbox[yaml]        # YAML policy files
pip install warderbox[all]         # Everything

Quick Start

from warderbox import WarderBox

shield = WarderBox(budget=5.00, auto_snapshot=True)

@shield.tool(cost=0.03)
def query_db(sql: str) -> str:
    return db.execute(sql)

# Normal execution
result = query_db(sql="SELECT * FROM users")
print(result.output)  # Works fine

# Destructive SQL is auto-blocked
result = query_db(sql="DROP TABLE users")
print(result.blocked)       # True
print(result.block_reason)  # "CRITICAL risk blocked: Destructive SQL detected"

# Undo the last action
shield.rollback()

Core Features

Policy-as-Code

Define rules in JSON or YAML files:

{
  "budget": 25.00,
  "mode": "production",
  "global_rpm": 120,
  "tool_limits": {"send_email": {"rpm": 10, "rph": 100}},
  "rules": [
    {
      "name": "no_prod_access",
      "match": {"args_regex": "production|prod-db"},
      "risk": "critical",
      "reason": "Production database access blocked"
    }
  ],
  "webhooks": [
    {"url": "https://hooks.slack.com/...", "events": ["action_blocked"]}
  ]
}

from warderbox import load_policy
shield = load_policy("policy.json")

Multi-Agent Support

from warderbox import AgentPool

pool = AgentPool(default_budget=5.00)
researcher = pool.register("researcher", budget=3.00)
coder = pool.register("coder", budget=10.00)
reviewer = pool.register("reviewer", budget=1.00, mode="test")

pool.stats()        # Aggregated view across all agents
pool.over_budget()  # ["researcher"] if over budget

Rate Limiting

from warderbox import WarderBox, ToolLimit

shield = WarderBox(
    budget=10.00,
    global_rpm=60,                                    # 60 calls/min globally
    tool_limits={"send_email": ToolLimit(rpm=5)},     # 5 emails/min
)

Retry with Backoff

from warderbox import RetryPolicy

policy = RetryPolicy(max_retries=3, backoff_base=1.0, jitter=0.5)

@shield.tool(cost=0.05, retry=policy)
def flaky_api(query: str) -> str:
    return requests.get(f"https://api.example.com?q={query}").text

Sensitive Data Masking

from warderbox import DataMasker

masker = DataMasker()
shield.set_masker(masker)

# Emails, API keys, credit cards, SSNs, JWTs are auto-masked in logs
masker.add_pattern("project_id", r"PROJ-\d{6}", "PROJ-******")

Enterprise Safety

Circuit Breaker

Auto-disable tools that fail repeatedly (CLOSED -> OPEN -> HALF_OPEN):

from warderbox import CircuitBreaker

cb = CircuitBreaker(failure_threshold=5, reset_timeout=60.0)
shield.set_circuit_breaker(cb)
# After 5 consecutive failures, the tool is auto-disabled for 60s
# Then one probe call is allowed (HALF_OPEN) to test recovery

Human-in-the-Loop Approval

from warderbox import ApprovalManager

approver = ApprovalManager(default_timeout=30.0)
approver.require_approval("send_email", timeout=60.0)
approver.require_approval_for_risk("HIGH")
shield.set_approval(approver)

# From another thread or webhook handler:
approver.approve(request_id, approver="admin@company.com")

Agent Capabilities / Permissions

Fine-grained tool access control per agent:

from warderbox import CapabilityManager

caps = CapabilityManager()
caps.grant("researcher", allow=["search", "read_file"])
caps.grant("intern_bot", deny=["delete_record", "send_email"])
caps.add_constraint("coder", "run_sql", args_deny=["DROP", "TRUNCATE"])
shield.set_capabilities(caps)

Compliance & Audit Reports

from warderbox import ComplianceReporter

reporter = ComplianceReporter(db_path="audit.db")
report = reporter.generate_report(start="2026-01-01", end="2026-12-31")
print(report.summary)          # Text summary with anomaly detection
reporter.export_json("audit.json")
html = reporter.to_html()      # Full HTML report

Observability

Distributed Tracing

OpenTelemetry-compatible span system:

from warderbox import Tracer

tracer = Tracer(service_name="my-agent")
shield.set_tracer(tracer)

# Every tool call automatically generates spans
# Nested spans for complex workflows:
with tracer.span("data_pipeline") as parent:
    with tracer.span("extract", parent=parent):
        extract_data()
    with tracer.span("transform", parent=parent):
        transform_data()

tracer.export_json("traces.json")  # JSON export
tracer.export_otlp()               # OTLP-compatible dict

Cost Forecasting

from warderbox import CostForecaster

forecaster = CostForecaster(shield)
report = forecaster.forecast()
print(report.burn_rate_per_hour)       # $/hour
print(report.estimated_exhaustion)     # datetime
print(report.estimated_actions_remaining)
print(report.trend)                    # "increasing" / "stable" / "decreasing"

Agent Scoring & Benchmarking

from warderbox import AgentScorer, PoolScorer

scorer = AgentScorer(shield, agent_name="my-bot")
report = scorer.score()
print(report.overall_score)    # 0-100
print(report.grade)            # A/B/C/D/F
print(report.safety_score)     # Safety dimension
print(report.reliability_score)

# Rank all agents in a pool
pool_scorer = PoolScorer(pool)
print(pool_scorer.summary())   # Ranked leaderboard

Web Dashboard

warderbox --db audit.db dashboard --port 8080

Real-time web UI with 4 tabs: Overview, FinOps, Routing, and Alerts.

Smart Model Routing (FinOps)

from warderbox import ModelRouter

router = ModelRouter()
router.add_model("haiku", cost_per_1k=0.00025, max_complexity=0.35)
router.add_model("sonnet", cost_per_1k=0.003, max_complexity=0.65)
router.add_model("opus", cost_per_1k=0.015, max_complexity=1.0)

decision = router.route("Summarize this text")
print(decision.model)           # "haiku" (cheapest capable)
print(decision.estimated_cost)  # $0.00025/1k tokens

stats = router.stats()
print(stats.savings_percent)    # 60-90% cost savings

Cost Anomaly Detection

from warderbox import AnomalyDetector

detector = AnomalyDetector(sensitivity=2.0, loop_threshold=10)

# Records costs and fires alerts on spikes
detector.record("search_api", 0.03)
detector.record("search_api", 5.00)  # spike! returns alerts

# Alert types: COST_SPIKE, RATE_SPIKE, BUDGET_BURN, LOOP_DETECTED
print(detector.stats())

Prometheus Metrics

from warderbox.metrics import start_metrics_server
start_metrics_server(shield, port=9090)
# Scrape http://localhost:9090/metrics

Exports: warderbox_actions_total, warderbox_actions_blocked_total, warderbox_budget_spent_dollars, per-tool and per-agent labels.

Webhook Notifications

shield.notifier.add_webhook(
    url="https://hooks.slack.com/services/xxx",
    events={"action_blocked", "budget_exceeded"},
)

shield.notifier.on("action_blocked", lambda event, data: print(f"Alert: {data}"))

Resilience Testing

Chaos Engineering

Test agent resilience with controlled fault injection:

from warderbox import ChaosEngine, FaultType

chaos = ChaosEngine()
shield.set_chaos(chaos)

# Random failures (20% of calls)
chaos.add_fault("search_api", FaultType.ERROR, probability=0.2)

# Latency injection
chaos.add_fault("slow_api", FaultType.LATENCY, latency_ms=2000)

# Conditional faults
chaos.add_fault("db_query", FaultType.ERROR,
                condition=lambda args: "prod" in str(args))

# Wildcard — affect all tools
chaos.add_fault("*", FaultType.LATENCY, latency_ms=500, probability=0.1)

Action Replay

from warderbox.replay import ActionReplayer

replayer = ActionReplayer(db_path="audit.db", executors={"run_sql": my_fn})
report = replayer.replay_all(shield)
print(f"Match rate: {report.success_rate:.0%}")

Async Support

from warderbox import AsyncWarderBox

shield = AsyncWarderBox(budget=5.00)

@shield.async_tool(cost=0.05)
async def async_query(sql: str) -> str:
    return await db.execute(sql)

Framework Integrations

# OpenAI
from warderbox.integrations.openai_agents import OpenAIShieldMiddleware
middleware = OpenAIShieldMiddleware(shield=shield, executors={...})

# Anthropic
from warderbox.integrations.anthropic_sdk import AnthropicShieldMiddleware
middleware = AnthropicShieldMiddleware(shield=shield, executors={...})

# LangChain
from warderbox.integrations.langchain import shield_langchain_tools

# CrewAI
from warderbox.integrations.crewai import shield_crewai_tools

# AutoGen
from warderbox.integrations.autogen import shield_autogen_functions

CLI

warderbox stats                    # Action statistics
warderbox logs                     # View action log
warderbox logs --blocked           # Only blocked actions
warderbox export audit.json        # Export to JSON
warderbox validate policy.json     # Validate a policy file
warderbox dashboard --port 8080    # Start web dashboard

Risk Levels

Architecture

Agent -> WarderBox -> Tool
         |-- Capability Check (permissions)
         |-- Circuit Breaker (fail-fast)
         |-- Risk Classification (rules engine)
         |-- Approval Workflow (human-in-the-loop)
         |-- Budget Check + Forecast
         |-- Rate Limit Check
         |-- Snapshot (for rollback)
         |-- Chaos Injection (testing)
         |-- Execute (or simulate in test mode)
         |-- Retry (on failure, with backoff)
         |-- Data Masking (for logs)
         |-- Log to SQLite
         |-- Emit webhook/callback events
         |-- Record span (distributed tracing)
         |-- Agent Scoring
         |-- Smart Model Routing (FinOps)
         |-- Cost Anomaly Detection
         |-- Prometheus metrics

Stats

• 36 Python modules, 255+ tests
• Zero mandatory dependencies
• 5 framework integrations
• Works with Python 3.11+

Author

Ali Celik — LinkedIn · Email

License

MIT — see LICENSE for details.