@computegauge/mcp

The cost intelligence + agent credibility layer for AI agents. Install once, every session is cost-aware and credibility-building.

npm install -g @computegauge/mcp

Why This Exists

AI agents are powerful but expensive. A single Claude Code session can cost $5-50+. A multi-agent workflow can burn through $100 in minutes. And nobody tells the agent — or the human — until the bill arrives.

ComputeGauge MCP makes agents cost-aware AND gives them a credibility score. Any MCP-compatible agent (Claude, Cursor, Windsurf, custom agents) gets:

1. Cost Intelligence — Know which model is optimal for every task, track spend in real-time
2. Agent Credibility — Build a reputation score (0-1000) by making smart decisions. Compete on a leaderboard.
3. Local→Cloud Routing — Detect when local inference isn't good enough, route to cloud, earn credibility for smart routing

The result: agents that spend 40-70% less, build visible credibility, and know when to route to cloud.

Setup — 30 Seconds

Claude Desktop / Claude Code

Add to ~/.config/claude/claude_desktop_config.json:

{
  "mcpServers": {
    "computegauge": {
      "command": "npx",
      "args": ["-y", "@computegauge/mcp"]
    }
  }
}

That's it. Restart Claude. Every conversation now has cost intelligence + credibility tracking.

With Provider API Keys (Enhanced)

{
  "mcpServers": {
    "computegauge": {
      "command": "npx",
      "args": ["-y", "@computegauge/mcp"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "OPENAI_API_KEY": "sk-...",
        "COMPUTEGAUGE_BUDGET_TOTAL": "50"
      }
    }
  }
}

With Local Inference (Ollama, vLLM, etc.)

{
  "mcpServers": {
    "computegauge": {
      "command": "npx",
      "args": ["-y", "@computegauge/mcp"],
      "env": {
        "OLLAMA_HOST": "http://localhost:11434",
        "OLLAMA_MODELS": "llama3.3:70b,qwen2.5:7b,deepseek-r1:14b",
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "COMPUTEGAUGE_BUDGET_TOTAL": "50"
      }
    }
  }
}

Cursor

Add to Cursor MCP settings:

{
  "computegauge": {
    "command": "npx",
    "args": ["-y", "@computegauge/mcp"]
  }
}

Tools Reference

Agent-Native Tools (use automatically every session)

Credibility Tools (the reputation protocol)

Intelligence Tools (for user questions)

Resources

Prompts

Agent Credibility System

Every smart decision earns credibility points on a 0-1000 scale:

Credibility Tiers

Earnable Badges

Local Cluster Integration

ComputeGauge auto-detects local inference endpoints:

Set OLLAMA_MODELS="llama3.3:70b,qwen2.5:7b" (comma-separated) to declare available models.

The Local→Cloud Routing Flow

1. Agent calls assess_routing("code_generation", quality="good")
2. ComputeGauge checks: local llama3.3:70b quality for code_generation = 80/100
3. "Good" quality threshold = 78 → Local model is sufficient!
4. Agent uses local model → saves money → earns credibility for honest assessment

OR:

1. Agent calls assess_routing("complex_reasoning", quality="excellent")
2. ComputeGauge checks: local llama3.3:70b quality for complex_reasoning = 78/100
3. "Excellent" quality threshold = 88 → Quality gap of 10 points → Route to cloud!
4. Agent calls pick_model → gets Claude Sonnet 4 → executes → calls route_to_cloud
5. Agent earns +70 credibility points for smart routing decision

How pick_model Works

The decision engine scores every model across three dimensions:

Quality — Per-task-type scores for 14 task types Cost — Real pricing from 8 providers, 20+ models, calculated per-call (log-scale normalization) Speed — Relative inference speed scores

Model Coverage

Local Models Supported

Environment Variables

For Agent Developers

If you're building AI agents (via Claude Agent SDK, LangChain, CrewAI, AutoGen, etc.), ComputeGauge MCP is the easiest way to add cost awareness AND agent credibility:

1. Zero integration effort — Just add the MCP server to your agent's config
2. No code changes — The agent discovers 18 tools via MCP protocol automatically
3. Immediate value — pick_model returns recommendations on first call, credibility tracking starts automatically
4. Session tracking built-in — Full cost visibility per agent run
5. Credibility system — Your agent earns a visible reputation score that users can see
6. Local cluster support — Auto-detect and leverage on-prem inference
7. Budget guardrails — Warnings when approaching limits

Pattern: Cost-Aware + Credibility-Building Agent Loop

1. Agent receives task
2. Agent calls assess_routing(task_type) → local or cloud?
3. Agent calls pick_model(task_type, priority="balanced")
4. Agent uses recommended model for the task
5. Agent calls log_request(provider, model, tokens)
6. Agent calls rate_recommendation(model, rating, success)
7. If cloud-routed: agent calls route_to_cloud(task_type, reason, model)
8. Every 5 requests, agent calls session_cost()
9. If session cost > 80% of budget, switch to priority="cheapest"
10. At session end: check credibility_profile()

This pattern reduces costs by 40-70% while building a credibility score that makes users trust the agent more.

License

Apache-2.0 — Free to use, modify, and distribute.

Links

• ComputeGauge Dashboard
• CLI (@computegauge/cli)
• Adapters (@computegauge/adapters)