Lutflow MCP Server

The first AI FinOps tool with native MCP support.

Works inside Cursor, Claude Code, VS Code, and any MCP-compatible client.

Installation

pip install lutflow-mcp

Or install from source:

cd mcp-server
pip install -e .

Configuration

For Cursor

Add to your ~/.cursor/mcp.json:

{
  "mcpServers": {
    "lutflow": {
      "command": "lutflow-mcp",
      "env": {
        "LUTFLOW_API_KEY": "lut_xxx",
        "LUTFLOW_TENANT_ID": "your-tenant"
      }
    }
  }
}

For Claude Desktop

Add to your Claude Desktop config:

{
  "mcpServers": {
    "lutflow": {
      "command": "lutflow-mcp",
      "env": {
        "LUTFLOW_API_KEY": "lut_xxx"
      }
    }
  }
}

Available Tools

1. check_budget

Check current budget status for a tenant.

check_budget(tenant_id="acme")

Returns: budget limit, spent amount, remaining budget, utilization percentage.

2. list_deployments

List all active AI model deployments.

list_deployments(tenant_id="acme")

Returns: deployment IDs, models, costs, status, kill path state.

3. get_costs

Get detailed cost breakdown by model and deployment.

get_costs(period="today")
get_costs(deployment_id="abc-123", period="week")

Returns: total spend, per-model costs, billing event count.

4. kill_deployment

Manually terminate an AI model deployment.

kill_deployment(deployment_id="abc-123", reason="overspend")

Triggers the Lutflow kill path (eBPF → Tetragon → kubectl fallback).

5. recommend_model

Get AI model recommendation for a task within budget.

recommend_model(task="text-classification", budget_usd_per_hour=0.50)
recommend_model(task="code-generation", gpu_type="nvidia-a100-80gb")

Uses live HuggingFace benchmark data and real-time GPU pricing.

Environment Variables

Variable	Description	Default
LUTFLOW_API_KEY	API key for Lutflow Cloud	None (offline mode)
LUTFLOW_ENDPOINT	API endpoint	https://api.lutflow.dev
LUTFLOW_TENANT_ID	Default tenant ID	default

Offline Mode

When the Lutflow API is unreachable, the MCP server operates in offline mode:

• check_budget returns default/cached values
• list_deployments returns empty list
• get_costs returns zero costs
• kill_deployment cannot confirm termination
• recommend_model uses cached recommendations

All responses in offline mode are clearly labeled [OFFLINE MODE].

Example Usage in Cursor

Once configured, you can ask Claude:

"Check my AI budget status"

"List all my active deployments"

"How much did I spend on GPT-4 today?"

"Kill deployment abc-123 — it's overspending"

"Recommend a model for text classification under $0.50/hr"

Links

• Website: lutflow.dev
• SDK: pypi.org/project/lutflow
• GitHub: github.com/Lutflow/lutflow

License

Business Source License 1.1 (BSL 1.1)