Overllm

Find the LLM call you didn't need — your GPT call is a regex. Static, zero-config linter that flags unnecessary GPT/Claude/AI API calls a regex, the stdlib, or a library already does (parsing dates, extracting emails/JSON, sorting, classifying with fixed labels). Runs locally, no model, no API key, nothing phones home. Python + JS/TS. Observability tells you what you spent; overllm finds the call you never should have made — cutting LLM/API cost as a result, not a promise.

overllm — find the LLM call you didn't need

Your GPT call is a regex. Use this skill to audit a codebase for wasteful or unnecessary GPT/Claude/LLM API calls — the ones where you pay latency, money, and nondeterminism for something a regex, the stdlib, or a library already does (parsing a date, extracting an email or JSON, sorting a list, classifying with a fixed set of labels). overllm finds the model calls you can delete outright; cutting your LLM/API bill is the result, not the pitch.

Cost dashboards (Helicone, Langfuse) tell you what you already spent, and caches make a call cheaper. Neither asks whether you needed the call. overllm does — from the source, before anything runs.

overllm is a static, zero-config linter. It parses code locally (Python via the stdlib ast, JavaScript/TypeScript via tree-sitter), runs no model, makes no network call, needs no API key, and sends no telemetry — nothing phones home. Same code in, same result out.

Your GPT call is a regex — see it

python

# before — a paid, nondeterministic API round-trip to pull an email out of text
email = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": f"extract the email from: {text}"}],
).choices[0].message.content

# after — free, instant, deterministic
import re
email = re.search(r"[\w.+-]+@[\w-]+\.[\w.-]+", text)

overllm flags the first as llm-extraction and points you straight at the fix. It is not an "AI slop" / AI-generated-code detector — it flags wasteful AI usage (calls that shouldn't exist), not AI-written code smell.

When to use

• "How do I reduce my OpenAI/Claude API bill / cut token usage?"
• "Audit this repo for unnecessary GPT/Claude/LLM calls."
• "Where am I calling a model to do something plain code could do?"
• "Check this file before I commit — am I wasting an API call here?"
• Pre-commit / CI gating on needless AI calls.

How to run

Check the binary is available first; if missing, install it:

bash

overllm --version || pip install "overllm[js]"

overllm[js] adds JavaScript/TypeScript support; plain pip install overllm covers Python only.

Then scan. Always pass --format json and --exit-zero so you get structured findings and overllm's non-zero "findings exist" exit code is not mistaken for a crash:

bash

overllm <path> --format json --exit-zero

• One file: overllm app.py --format json --exit-zero
• A folder: overllm src/ --format json --exit-zero
• Whole project: overllm . --format json --exit-zero

Useful flags:

• --all — also surface the static-prompt info-level rule (quiet by default).
• --min-severity {error,warning,info} — default shows warning and above.
• --select <ids> / --ignore <ids> — comma-separated rule ids to run or skip.
• --format {human,json,sarif,markdown,github} — use json to reason over findings, markdown for a pasteable report, github for CI annotations.
• --baseline / --write-baseline — adopt on an existing repo: snapshot current findings, then report only new ones.
• --fix (safe) / --fix --unsafe-fixes / --diff — auto-remove a rejected sampling param, or (unsafe) swap a retired model id; --diff previews.

How to present findings

Each finding has a file, line, column, rule id, a message, and a concrete deterministic replacement (-> use sorted(), -> batch the inputs into a single call). When you report back:

1. Lead with the count: "N needless LLM calls across M files."
2. For each finding, give file:line, what it's doing wrong in one line, and the suggested fix. The fix is the value — always include it.
3. If overllm prints nothing, say so plainly: a clean codebase exits 0 with no findings. That's a pass, not a failure.

overllm is quiet by design and only fires on concrete patterns with a named replacement — if it floods the user, treat that as a bug worth reporting, not signal.

What it does not do

It does not lint the code an AI wrote and never runs or calls a model. By default it only reads and reports — a plain scan changes nothing, worst case a few lines of output. The one exception is the opt-in --fix flag (see Useful flags above): it edits files, and only for the two mechanically-safe fixes (re-parsed before writing, so it never breaks a file). A scan without --fix is always read-only.