mcp-tool-chain-optimizer

MCP server that makes multi-step AI agent tool chains more reliable.

Analyze any sequence of tools → get success probability, find the bottleneck, see a better execution order, and receive concrete improvement tips.
Everything runs locally with pure math – zero external API calls, zero extra cost.

Built for entrepreneurs and AI builders who want accountable, predictable agents (part of the Prevalid AI Execution OS vision).

Why this exists

When an AI agent chains 5–10 tools together, small failure rates multiply:

• 90% × 85% × 92% × 80% ≈ 56% overall success
• One weak “critical” tool can silently kill the whole workflow
• Cost and latency explode without anyone noticing

This MCP server gives the agent (or the human developer) a fast, free way to measure and improve that chain before it goes to production.

Tools

Tool	What it does
analyze_tool_chain	Full report: probability, risk level, cost, latency, bottleneck, suggestions, better order
estimate_chain_success	Quick probability from a simple list of success rates
find_bottlenecks	Rank the weakest links (success rate × impact)
suggest_better_order	Fail-fast reordering that still respects dependencies
generate_reliability_report	Human-readable Markdown report ready to share with stakeholders

Quick Start

# Install
pip install -e .

# Run the MCP server (stdio)
mcp-tool-chain-optimizer
# or
python -m mcp_tool_chain_optimizer.server

Claude Desktop / Cursor / any MCP client

Add to your MCP config:

{
  "mcpServers": {
    "tool-chain-optimizer": {
      "command": "python",
      "args": ["-m", "mcp_tool_chain_optimizer.server"],
      "cwd": "/path/to/mcp-tool-chain-optimizer"
    }
  }
}

Example

[
  {"name": "web_search", "success_rate": 0.92, "avg_latency_ms": 800, "cost_per_call": 0.002, "failure_impact": "medium"},
  {"name": "extract_entities", "success_rate": 0.78, "avg_latency_ms": 300, "cost_per_call": 0.001, "failure_impact": "high"},
  {"name": "write_summary", "success_rate": 0.95, "avg_latency_ms": 1200, "cost_per_call": 0.005, "failure_impact": "low", "depends_on": ["extract_entities"]}
]

→ Overall success ≈ 68%, bottleneck = extract_entities, suggested order puts the risky extractor earlier (fail-fast).

Design Principles

• Type A (mcpize): pure computation, zero paid API
• Local-first, privacy-friendly
• Fast enough for real-time agent self-reflection
• Simple JSON in / Markdown out – works with any LLM

Development

pip install -e ".[dev]"
pytest

License

MIT

---

Made with ❤️ for the Prevalid community – making AI agents accountable at the infrastructure level.