mesh-cost-firewall

Real-time AI agent cost gating system that prevents budget overruns before they happen.

What is this?

mesh-cost-firewall is a cost control middleware that sits between your agent mesh and LLM APIs, automatically rejecting or rate-limiting requests that would breach per-agent, per-user, or global budget limits. Instead of discovering runaway costs in your billing dashboard, agents hit configurable spending ceilings and fail gracefully. Built for autonomous multi-agent systems where financial guardrails are essential.

Features

• Multi-level budget enforcement — Global, per-user, and per-agent cost caps
• Real-time cost estimation — Pre-request cost prediction before API calls
• Graceful rate limiting — Circuit breaker pattern for budget exhaustion
• LLM model awareness — Built-in pricing for GPT-4, Claude, and other major models
• Request interception — Transparent proxy layer between agents and APIs
• Configuration as code — YAML/env-based budget definitions
• Docker-native — Ships with Dockerfile and docker-compose for instant deployment
• Production-ready — Handles concurrent requests, persistent state, and monitoring hooks

Quick Start

Installation

# Clone and install
git clone <repo>
cd mesh-cost-firewall
pip install -e .

# Or with Docker
docker-compose up -d

Configuration

Create a .env file (see .env.example):

GLOBAL_BUDGET_CENTS=100000
DEFAULT_USER_BUDGET_CENTS=10000
DEFAULT_AGENT_BUDGET_CENTS=1000

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

Basic Usage

from mesh_cost_firewall import CostFirewall

firewall = CostFirewall(config_path="config.yaml")

# Wrap your LLM call
result = firewall.request(
    agent_id="researcher_1",
    user_id="user_42",
    model="gpt-4",
    tokens=2000,
    endpoint="https://api.openai.com/v1/chat/completions"
)

if result.allowed:
    # Safe to proceed
    response = call_llm_api(...)
else:
    # Budget exhausted
    logger.warning(f"Request rejected: {result.reason}")

Or use as an HTTP proxy:

export FIREWALL_PORT=8000
python -m mesh_cost_firewall.main
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "X-Agent-Id: researcher_1" \
  -H "X-User-Id: user_42" \
  -d '{"model": "gpt-4", ...}'

Tech Stack

• Python 3.9+ — Core runtime
• FastAPI — HTTP proxy and API
• Pydantic — Configuration validation
• APScheduler — Budget reset scheduling
• Docker — Containerization
• SQLite — Persistent state (upgradeable to Postgres)

License

MIT