LLM Checker

What it does

LLM Checker is a Node.js command-line tool that inspects the machine it runs on (CPU, GPU, RAM and acceleration backend) and ranks local language models by how well they will fit and run there. The same npm package ships a second program, `llm-checker-mcp`, which is a stdio MCP server that puts that analysis, and control of a local Ollama install, in front of an assistant.

Through the server an assistant can detect hardware, get ranked or per-category model suggestions, search the bundled model catalog, plan GPU placement and Ollama context or parallel settings, and check a GGUF or safetensors file structurally before loading it (the ModelVet WebAssembly verifier, bundled). It can also act on Ollama: list, pull, run prompts against, benchmark, compare and delete models, and read live memory use of loaded models. `project_recommend` scans a project folder to suggest a model suited to its languages, and `cli_exec` runs any of 24 allowlisted `llm-checker` CLI subcommands, among them `ai-run` and `sync`.

The v3.8.2 server file registers 28 tools, the same 28 the README tables list. The project folder scan stops at five levels deep and skips dotfolders, node_modules, dist, build, target, vendor and __pycache__; `cli_exec` refuses more than 100 arguments and times out after 180 seconds unless `timeout_ms` (1 to 600 seconds) is given (bin/mcp-server.mjs at v3.8.2).

What it needs

Node.js 18 or newer. Ollama installed and running for the model-management, run and benchmark tools; `ollama_pull` calls the `ollama` program, so it must be on PATH, and the HTTP calls go to `OLLAMA_HOST` (http://localhost:11434 by default). The hardware and recommendation tools work without Ollama. SQLite-backed commands need the optional `sql.js` dependency; reinstall with `npm install -g llm-checker --include=optional` if a package manager skipped it.

Limits

It only manages models through Ollama; other runtimes (vLLM, MLX, llama.cpp, transformers) appear as recommendation targets, not as things it can pull or run. Recommendation and auto-selection exclude models labelled uncensored, abliterated or heretic unless `--include-uncensored` is passed. `ollama_pull` gives up after 10 minutes. `verify_model` checks file structure only, not what the model will say. The scoped GitHub Packages mirror `@pavelevich/llm-checker` may lag behind npm; install `llm-checker` from npm.

Who it suits

Suits someone running models locally with Ollama who wants an assistant to answer which model fits this machine, then pull, benchmark or remove models without leaving the chat. It runs fully on your own machine and needs no account. Because tools can download multi-gigabyte models, delete installed ones and run allowlisted CLI commands, pick a client that asks before each tool call if you want to approve those steps. If you only use hosted model APIs and no local inference, it has little to offer.