Welcome to local-llm-connector-mcp

[lefos13]

🚀 Announcing v1.0.0 of local-llm-connector-mcp: Stop letting massive test logs bloat your AI agent's context!

Hey everyone! 👋

I'm excited to announce the first release (v1.0.0) of local-llm-connector-mcp (available on GitHub at lefos13/local-llm-connector-mcp)!

If you use AI coding agents (like Codex, Claude Desktop, or Antigravity) to write code and run test suites, you’ve probably run into agent context bloat. An agent runs npm test or pytest, gets 5,000 lines of noisy compiler warnings or stack traces, and suddenly your context window is flooded. This wastes tokens, slows down responses, and jacks up API costs.

local-llm-connector-mcp solves this by offloading noisy logs and workspace exploration to a local LLM (like Qwen2.5-Coder running on Ollama or llama.cpp). The server runs validation commands in your workspace, saves full logs to disk, and feeds a smart, trimmed context to your local LLM. It then returns a compact, structured JSON verdict back to your main agent.

---

🛠️ Key Features

• run_test_verdict: Runs explicit or auto-detected validation commands (supports Node.js, Rust, Go, Python). Returns a structured JSON verdict (pass, fail, uncertain) and optional autoTriage to attach root-cause analyses on failure.
• run_command_digest: Generalizes log digestion for other noisy commands like npm install, container builds, database migrations, git log, etc., using an "intent" prompt to steer the local summary.
• scout_codebase: A local recon tool. It greps the codebase deterministically, then uses the local LLM to rank candidate files and line ranges so your main agent knows exactly where to read without doing a blind search.
• run_changed_files_review: Light-weight, Git-diff-aware review of changed files (under 500 KB) before running expensive validation runs.
• query_log & grep_log: Interrogate or search saved log files on disk (by a stable runId) without re-reading the raw files into the main agent's context.
• 📈 Context Savings Dashboard: Run npm run analytics:ui to see a local dashboard (http://127.0.0.1:8787) tracking how many tokens you've saved, tool latency, and overall cost reduction.

Tip

It runs beautifully with the Qwen2.5-Coder-7B-Instruct model (e.g. Q4_K_M GGUF variation) using a Macbook M-series chip. The right model is the one which is right for your hardware. The better hardware the better the model, the better the results of this tool.

---

🚀 Quick Start

1. Install & Build

git clone https://github.com/lefos13/local-llm-connector-mcp.git
cd local-llm-connector-mcp
npm install
npm run build

2. Start a Local LLM

Ensure you have a local OpenAI-compatible endpoint running (e.g., using Ollama or Llama.cpp).
For Ollama:

ollama run qwen2.5-coder:7b-instruct

3. Add to your MCP Client Configuration

Add the server to your client's config file (e.g. Claude Desktop Config or Codex or Antigravity 2.0 or whatever supports mcp servers):

{
  "mcpServers": {
    "local-tester": {
      "command": "node",
      "args": ["/absolute/path/to/local-tester-mcp/dist/index.js"],
      "env": {
        "LOCAL_LLM_API_URL": "http://localhost:11434/v1",
        "LOCAL_LLM_MODEL": "qwen2.5-coder:7b-instruct"
      }
    }
  }
}

---

💬 We Need Your Feedback!

Since this is the first release, I would love to get your thoughts on how to improve the tool. If you decide to give it a spin, please let me know:

1. Local LLM Performance: Which local model/size did you use, and how accurate were the test verdicts or triage suggestions?
2. Tool Set & UX: Are the tool interfaces (like run_test_verdict and scout_codebase) intuitive for your coding workflows?
3. Analytics: Did you check out npm run analytics:ui? What other metrics or insights would you like to see tracked?
4. Bugs & Edge Cases: Did command auto-detection fail for your specific Python, Go, Rust, or Node project structure?

Feel free to post your thoughts, ideas, or screenshots of your local analytics dashboard below. If you run into bugs, please open a Github issue!

Thank you so much for testing! 🙏