AgentMolt — Kill Switch & Budget Control for AI Agents

Monitor, budget-cap, and kill-switch your AI agents. Works standalone or with OpenAI/Anthropic/LangChain/CrewAI.

→ agentmolt.dev

Install

pip install agentmolt

Quick Start — Try It Now (No API Key Needed)

from agentmolt import ACP, AgentBudget

# Create a local control panel (no API key = local-only mode)
panel = ACP()

# Register an agent with a $5 budget
panel.register("my-agent", budget=AgentBudget(max_cost_usd=5.00))

# Simulate agent making API calls
for i in range(10):
    try:
        state = panel.log("my-agent", model="gpt-4o", tokens=1000, cost=0.80)
        print(f"Call {i+1}: ${state.total_cost_usd:.2f} spent, alive={panel.is_alive('my-agent')}")
    except Exception as e:
        print(f"Call {i+1}: KILLED — {e}")
        break

Output:

Call 1: $0.80 spent, alive=True
Call 2: $1.60 spent, alive=True
Call 3: $2.40 spent, alive=True
Call 4: $3.20 spent, alive=True
Call 5: $4.00 spent, alive=True
Call 6: $4.80 spent, alive=True
Call 7: KILLED — Agent 'my-agent' killed: cost $5.60 exceeded budget $5.00

The agent is automatically killed when it exceeds its budget. No more runaway costs.

Kill Switch

from agentmolt import ACP

panel = ACP()
panel.register("research-bot")

# Check if agent is alive
print(panel.is_alive("research-bot"))  # True

# Kill it manually
panel.kill("research-bot", reason="suspicious activity")
print(panel.is_alive("research-bot"))  # False

# Any future calls raise AgentKilledException
try:
    panel.log("research-bot", model="gpt-4o", tokens=100, cost=0.01)
except Exception as e:
    print(f"Blocked: {e}")

# Revive if needed
panel.revive("research-bot")
print(panel.is_alive("research-bot"))  # True

Budget Enforcement

Set limits on cost, tokens, or number of calls:

from agentmolt import ACP, AgentBudget

panel = ACP()

panel.register("writer-agent", budget=AgentBudget(
    max_cost_usd=10.00,   # Kill at $10 spent
    max_tokens=100_000,    # Kill at 100K tokens
    max_calls=50,          # Kill after 50 API calls
    alert_at_pct=0.8,      # Trigger alert at 80% of any limit
))

Callbacks — Get Notified

def on_kill(agent_id):
    print(f"🚨 ALERT: {agent_id} was killed!")

def on_alert(agent_id, pct):
    print(f"⚠️ WARNING: {agent_id} at {pct:.0%} of budget")

panel = ACP(
    on_kill=on_kill,
    on_alert=on_alert,
)

panel.register("spender", budget=AgentBudget(max_cost_usd=1.00, alert_at_pct=0.5))

panel.log("spender", model="gpt-4o", tokens=500, cost=0.30)
# → nothing

panel.log("spender", model="gpt-4o", tokens=500, cost=0.30)
# → ⚠️ WARNING: spender at 60% of budget

panel.log("spender", model="gpt-4o", tokens=500, cost=0.50)
# → 🚨 ALERT: spender was killed!
# → raises AgentKilledException

Auto-Patch OpenAI (Zero Code Changes)

from agentmolt import ACP
from openai import OpenAI

panel = ACP()
client = OpenAI()
panel.patch_openai(client)

# Use OpenAI as normal — all calls are automatically monitored
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}]
)

# Check what happened
for agent_id, state in panel.all_agents().items():
    print(f"{agent_id}: {state.total_calls} calls, ${state.total_cost_usd:.4f}")

Auto-Patch Anthropic

from agentmolt import ACP
import anthropic

panel = ACP()
client = anthropic.Anthropic()
panel.patch_anthropic(client)

# All Claude calls are now monitored automatically

Decorators

from agentmolt import monitor, budget, kill_switch

@budget("analyst", max_cost_usd=50.0)
@kill_switch("analyst")
@monitor("analyst")
def run_analysis():
    # Your agent code here — automatically monitored
    pass

CLI

# Scan current directory for AI agent patterns
acp scan

# Show status of monitored agents
acp status

# Kill a rogue agent
acp kill marketing-bot --reason "unauthorized API spend"

# Check version
acp version

Status Dashboard

Check all your agents at a glance:

panel = ACP()
panel.register("agent-a", budget=AgentBudget(max_cost_usd=10.0))
panel.register("agent-b", budget=AgentBudget(max_cost_usd=5.0))

panel.log("agent-a", model="gpt-4o", tokens=2000, cost=1.50)
panel.log("agent-b", model="claude-3", tokens=500, cost=0.30)

for agent_id, state in panel.all_agents().items():
    status = "🟢 alive" if not state.killed else "🔴 killed"
    print(f"{agent_id}: {status} | ${state.total_cost_usd:.2f} | {state.total_calls} calls")

agent-a: 🟢 alive | $1.50 | 1 calls
agent-b: 🟢 alive | $0.30 | 1 calls

Connect to AgentMolt Dashboard

For real-time monitoring, alerts, and team dashboards:

panel = ACP(api_key="your-api-key")  # Get key at agentmolt.dev

Events are batched and sent every 5 seconds. Works offline — events are buffered and retried on failure.

How It Works

1. Register agents with optional budgets
2. Log every LLM call (or auto-patch OpenAI/Anthropic)
3. Budget check on every call — kills agent if exceeded
4. Kill switch — manual or automatic, blocks all future calls
5. Events buffered and flushed to dashboard (if API key set)

No external dependencies required for local mode. Just pip install agentmolt and go.

Requirements

• Python 3.9+
• No required dependencies for local mode
• Optional: openai, anthropic for auto-patching
• Optional: httpx for dashboard sync (included)

License

MIT — built by AgentMolt