steerplane 0.4.0

Agent Control Plane for Autonomous Systems — Runtime guards, loop detection, cost limits, and telemetry for AI agents.

SteerPlane SDK

Agent Control Plane for Autonomous Systems

"Agents don't fail in the dark anymore."

Quick Start

pip install steerplane

Decorator API (Minimal Integration)

from steerplane import guard

@guard(max_cost_usd=10, max_steps=50)
def run_agent():
    agent.run()

Context Manager API (Full Control)

from steerplane import SteerPlane

sp = SteerPlane(agent_id="my_bot")

with sp.run(max_cost_usd=10) as run:
    run.log_step("query_db", tokens=380, cost=0.002)
    run.log_step("generate_response", tokens=1240, cost=0.008)

Features

• 🛡️ Policy Engine (New in v0.4.0) — Define strict allow/deny rules, rate limits (sliding windows), and human-in-the-loop approval workflows for agents.
• 🌉 AI Gateway Proxy (New in v0.4.0) — Zero-code integration. Point your default LangChain/OpenAI client to the SteerPlane gateway, and every LLM call gets token tracking, cost evaluation, and policy enforcement automatically.
• 🧩 First-Class LangChain & CrewAI Support — Drop in SteerPlaneCallbackHandler for total observability of your multi-agent networks without modifying agent chains.
• 🔄 Infinite Loop Detection — Automatically detects repeating agent behavior and breaks loops before they drain API budgets.
• 💰 Hard Cost Limits — Stop expensive agent runs the instant they cross a predefined USD ceiling. Tracks 25+ LLM models locally.
• 🚫 Step Limits — Cap maximum execution steps to prevent runway execution.
• 📊 Deep Telemetry — Full step-by-step execution tracking (tokens, latency, cost per step) synced instantly to the dashboard.

Advanced Usage: Policy Engine

from steerplane import guard

@guard(
    max_cost_usd=10, 
    denied_actions=["DROP TABLE*", "rm -rf *"],
    rate_limits=[{"pattern": "send_email", "max_count": 5, "window_seconds": 60}],
    require_approval=["execute_trade*"]
)
def autonomous_agent():
    agent.run()

Documentation

See docs.steerplane.ai for full documentation.