🛡 agent-gov

AI Agent Cost Governance Platform

A reverse proxy that tracks, budgets, and controls what your AI agents spend. Like a credit card with limits — but for your agents.

What Problem Does This Solve?

AI agents call expensive tools (LLMs, browsers, APIs, email services). Without controls:

• A recursive agent burns ₹5,000 in one night on GPT-4
• A buggy loop sends 10,000 emails before you notice
• You discover the bill when your credit card statement arrives

agent-gov sits between your agents and their tools. Every call goes through us. We check budgets, track costs, and auto-pause overspending agents.

Quick Start

1. Install

pip install agent-gov-saas

2. Start the server

agent-gov start

Server starts at http://localhost:8000.

Alternatively, use Docker: docker compose up (see GitHub repo for docker-compose.yml).

3. Register an Agent

curl -X POST http://localhost:8000/agents/register \
  -H "Content-Type: application/json" \
  -d '{"name": "My Bot", "daily_budget": 500}'

Response:

{
  "api_key": "ag-abc123...",
  "name": "My Bot",
  "daily_budget": 500,
  "message": "Save this API key — it won't be shown again!"
}

4. Route Agent Calls Through the Proxy

Instead of calling tools directly, your agent calls our proxy:

# BEFORE (no governance):
response = call_openai(prompt)

# AFTER (with agent-gov):
gov_response = requests.post("http://localhost:8000/proxy/call", json={
    "agent_key": "ag-abc123...",
    "tool_name": "openai-gpt4",
    "estimated_cost": 12.50   # ₹12.50 for this call
})
if gov_response.status_code == 429:
    print("Budget exceeded! Agent auto-paused.")
else:
    response = call_openai(prompt)  # Proceed with actual call

5. Watch the Dashboard

Open http://localhost:8000/dashboard — live view of all agents, their spend, and budget status.

6. (Optional) Pre-Register Common Tools

# Register 24 common AI tools with realistic costs (₹)
python seed_tools.py

# Preview before registering:
python seed_tools.py --list

API Reference

Features

Architecture

Agent (your code)
    │
    │ POST /proxy/call { agent_key, tool_name, estimated_cost }
    ▼
┌──────────────────────────────┐
│         agent-gov             │
│                                │
│  1. Auth: Is key valid?       │
│  2. Budget: Will this exceed? │
│  3. Log: Track cost + tool    │
│                                │
│  If approved → return 200     │
│  If denied  → return 429     │
└──────────────────────────────┘
    │
    │ Agent calls actual tool
    ▼
┌──────────────────────────────┐
│   Actual Tool (OpenAI, etc.)  │
└──────────────────────────────┘

Technology Stack

License

MIT