Overllm MCP server

<!-- mcp-name: io.github.theadamdanielsson/overllm -->

overllm

Catch the LLM/AI calls you didn't need.

overllm is a small, fast linter with one job: find the places in your code where you call an AI model to do something plain code does better. You called GPT to parse a date. You called a model to extract JSON that json.loads already handles. You are paying latency, money, and nondeterminism for a regex.

It reads your code with a real parser: Python through the standard-library ast, and JavaScript and TypeScript through tree-sitter. No model runs, no network, no API key. Same code in, same result out. Fast enough for a pre-commit hook.

Cost dashboards and caches deal with calls you've already decided to make. overllm asks the earlier question — did you need the call at all — and answers it from the source, before you run anything. Everyone else lints the code the AI wrote; overllm catches where you're paying an AI to do what a library already does.

Install

pip install overllm          # Python
pip install "overllm[js]"    # adds JavaScript / TypeScript support

Use it

Point it at one file first, so you can see what it flags before you run it on everything:

overllm app.py       # one file
overllm src/         # a folder
overllm .            # the whole project

It reads your code and prints what it finds. By default it writes nothing and changes nothing — the worst case of a plain run is a few lines of output. The one mode that edits files is the opt-in --fix (below), and only for the two mechanically-safe fixes.

Example output:

app.py:42:5 llm-mechanical  LLM call asks the model to sort
    resp = client.chat.completions.create(model="gpt-4o", messages=[...])
    -> use sorted()

app.py:88:1 llm-in-loop  LLM call inside a loop: one API round-trip per iteration
    completion(model="gpt-4o", messages=[{"role": "user", "content": f"tag {x}"}])
    -> batch the inputs into a single call, cache repeated results, or use a function

2 needless LLM calls in 1 file.

It is quiet by default. Only warning and error findings show, so a clean project prints nothing and exits 0 — most codebases surface a handful or none. If it floods you, treat that as a bug and open an issue.

overllm exits non-zero when it finds something, so it gates a commit or a CI check. Pass --exit-zero to report without failing.

Your code stays on your machine

overllm is static analysis. It parses your files locally, then prints what it found. It never uploads your code, never calls an API, needs no key, and sends no telemetry — there is no model in the loop and nothing phones home. Pull your network cable and it runs exactly the same.

The core is a couple thousand lines of Python with no required dependencies, so you can read all of it before you trust it. overllm[js] adds tree-sitter to parse JavaScript and TypeScript; that is the only optional dependency.

Rules

Every rule fires only on a concrete code pattern, and every finding names the deterministic replacement. It stays silent when it is not sure.

By default overllm only raises warning and above, so it is quiet on your everyday code. static-prompt is info and stays silent unless you ask for it with --all or --min-severity info.

Shortened. The full README is on GitHub.