Streamline MCP

An MCP server to optimize LLM usage for coding agents. It provides tools to parse plans, estimate costs, recommend model strategies, and execute work units via LiteLLM.

Overview

Streamline helps you optimize your LLM usage by parsing plans, estimating costs across models, recommending the best strategy (Premium vs. Cheap vs. Mixed), and optionally executing the plan via LiteLLM.

Setup

1. Install Dependencies:

   npm install

2. Configure Models: Edit config/models.yaml to define your available models and pricing.

3. Configure Policies: Edit config/policies.yaml to set budget limits and risk tolerance.

4. Environment Variables: Create a .env file with:

   LITELLM_BASE_URL=http://localhost:4000
   LITELLM_API_KEY=sk-1234  # Optional if LiteLLM requires it

5. Build:

   npm run build

Usage via MCP

The server provides 4 main tools:

1. parse_plan: Converts markdown plan text into structured work units.
2. preflight_compare: Estimates token usage and cost for a given plan across suitable models.
3. recommend_strategy: Suggests "All Premium", "All Cheap", or "Mixed" strategies based on cost/performance trade-offs.
4. run_strategy: Executes a chosen strategy, logging costs and outcomes to a local SQLite database.

Running the Server

Run via node:

node dist/index.js

Or configure in your MCP client (Antigravity settings):

{
  "mcpServers": {
    "streamline": {
      "command": "npx",
      "args": ["-y", "streamline-mcp"],
      "env": {
        "LITELLM_BASE_URL": "http://localhost:4000"
      }
    }
  }
}

UI / Reporting

To visualize run history and cost metrics, generate a local HTML report:

npm run generate-report

This will create report.html in the root directory. Open it in any browser to see:

• Recent run statuses and costs
• Detailed step-by-step logs for the latest run
• Aggregate model success rates and efficiency metrics

Database

The server maintains a SQLite database model_optimizer.db in the working directory to track:

• Run history
• Cost vs Estimate accuracy
• Model success rates

Example Workflow

1. User provides a plan in Chat.
2. Antigravity calls parse_plan(plan_text).
3. Antigravity calls preflight_compare(plan_text, work_units).
4. Antigravity calls recommend_strategy(...).
5. User selects a strategy (e.g., "Mixed").
6. Antigravity calls run_strategy(strategy_id, allocations, ...) to execute.