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.

Rule	Severity	Fires when	Suggests
llm-mechanical	error	The prompt asks for a mechanical transform: sort, reverse, count, sum, deduplicate, change case, base64, arithmetic on literals.	the one-line stdlib equivalent
llm-extraction	error	The prompt asks the model to extract an email, URL, date, or number.	a regex, datetime, or urllib.parse
prompt-injection	error	Untrusted web-request input (request.args, request.json, ...) flows straight into the prompt.	keep it in a separate user message, validate it, constrain the model
llm-in-loop	warning	An LLM call runs once per loop iteration (real N calls, not streaming).	batch, cache, or move it out of the loop
deprecated-model	error / warning	The model id is a retired model (the call 404s) or one that is deprecated and scheduled for removal.	switch to the current model it names
unsupported-params	warning	temperature / top_p / top_k is set on a model that rejects them — the OpenAI reasoning (o1, o3, ...) series and the newest Anthropic models.	remove the parameter; steer with the prompt instead
json-mode-missing-json	error	response_format={"type": "json_object"} is set but the fully-static prompt never contains the word "json" — a guaranteed OpenAI 400.	add "json" to a message, or use a json_schema format
static-prompt	info	The user prompt is a compile-time constant, no variables. The input is fixed, so the call buys nothing.	precompute or cache the result

The last two check the call itself, not the prompt: a model id that no longer exists, or a knob the model ignores. Both are matched exactly against a known list, so a live model or alias is never flagged. The lists track provider deprecation pages and need updating over time.

It detects the OpenAI, Anthropic, Google, Mistral, Cohere, Groq, AWS Bedrock, HuggingFace, Replicate, LangChain, LiteLLM, and Ollama SDKs in Python, the Vercel AI SDK (generateText, streamText, generateObject) and the openai / anthropic node SDKs in JavaScript and TypeScript, and raw HTTP requests to those hosts. It also follows a model through LCEL composition — a chain = prompt | model | parser pipe, a bound model (.with_structured_output(...)), or an alias — so chain.invoke(...) is seen; embeddings calls (embeddings.create) count too. When a call goes through your own wrapper or a framework overllm can't see, name it in llm_calls (below).

Silence a false positive

resp = client.chat.completions.create(...)  # overllm: ignore
resp = client.chat.completions.create(...)  # overllm: ignore=llm-in-loop

Put # overllm: ignore-file at the top of a file to skip the whole file.

Configure

In pyproject.toml (Python 3.11+):

[tool.overllm]
ignore = ["llm-in-loop"]
exclude = ["examples/", "migrations/"]
llm_calls = ["myapp.llm.ask", "chat_service.complete"]

Or on the command line: --select, --ignore, --min-severity, --all, and --config PATH (exclude is config-only). Run overllm --help for the full list.

Teaching overllm your own wrapper

Most code doesn't call the SDK inline — it wraps it (def ask(prompt): client.chat.completions.create(...)). overllm follows that wrapper on its own when it lives in the same file. When the call goes through a framework, a provider layer, or a **kwargs splat, overllm can't see the SDK call, so tell it the wrapper's name in llm_calls. After that, calls to ask(...) are treated like LLM calls — it reads the prompt argument and runs the loop and cost rules. A name matches bare (ask), dotted (myapp.llm.ask), or that dotted path imported under its short name.

Adopt on an existing codebase (baseline)

Dropping overllm on an old repo gives you a wall of findings you'll never get through. Snapshot them once and have CI flag only what's new after that:

overllm . --write-baseline        # writes overllm-baseline.json — commit it
overllm . --baseline              # reports only findings new since the snapshot
overllm . --update-baseline       # ratchet: also drop entries you've since fixed

The snapshot keys each finding on rule + file + code + model, not the line number, so unrelated edits above it don't invalidate it. And it counts how many times each one shows up instead of just diffing totals, so a new bad call still trips the check even if you happened to delete an old one somewhere else.

Fix what's safe (--fix)

Two of the rules have one obvious fix, so overllm can just do it for you:

overllm . --fix                  # drop a sampling param the model rejects (safe)
overllm . --fix --unsafe-fixes   # also swap a retired model id for its replacement
overllm . --fix --diff           # print the patch, don't touch anything

Plain --fix only does the safe one (unsupported-params). Swapping a model id changes what your code actually does at runtime, so that's behind --unsafe-fixes. Fixes edit the syntax tree, not the raw text, so your comments and strings are left alone, and overllm re-parses the file before saving — if the edit would break it, it's dropped. The other five rules need a human call, so it never touches them.

Pre-commit hook

In .pre-commit-config.yaml:

repos:
  - repo: https://github.com/theadamdanielsson/overllm
    rev: v0.6.0
    hooks:
      - id: overllm

GitHub Action

overllm ships an Action that scans a pull request and leaves one grounded comment. It stays silent when there is nothing to say.

name: overllm
on:
  pull_request:

permissions:
  contents: read
  pull-requests: write

jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: theadamdanielsson/overllm@v1
        with:
          paths: "."

Other output formats

overllm --format json .      # machine-readable
overllm --format sarif .     # upload to GitHub code scanning
overllm --format github .    # GitHub Actions inline annotations
overllm --format markdown .  # the PR-comment body

GitHub code scanning (SARIF)

overllm can output SARIF, so findings show up in the Security tab and inline on the diff. It's free on public repos:

name: overllm-scan
on: [push, pull_request]
permissions:
  contents: read
  security-events: write   # required to upload results
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: pipx run overllm . --format sarif --exit-zero > overllm.sarif
      - uses: github/codeql-action/upload-sarif@v3
        with:
          sarif_file: overllm.sarif
          category: overllm

Use it from an agent (MCP server)

overllm ships an MCP server, so an AI agent (Claude Desktop, Cursor, or anything that speaks MCP) can call it to audit code for needless model calls. It exposes two tools: scan_path (a file or directory on disk) and scan_code (a snippet passed inline). Both return the same findings the CLI does — the rule, the location, and the concrete replacement — and neither writes or changes anything.

Install the server variant and point your client at it (the server needs Python 3.10+; the linter itself still runs on 3.9):

pip install "overllm[mcp]"

{
  "mcpServers": {
    "overllm": {
      "command": "uvx",
      "args": ["--from", "overllm[mcp]", "overllm-mcp"]
    }
  }
}

Then ask the agent things like "scan this repo for unnecessary LLM calls" or "is this function wasting a model call?" and it gets grounded, deterministic findings instead of guessing.

Why not just use an AI code reviewer?

AI reviewers and AI-slop linters look at the code the model produced: comments, dead code, structure. None of them ask the question overllm asks, which is whether you needed the model at all. It is a different axis, and it is one plain static analysis can answer with high precision and zero cost.

overloop, for your running agent

overllm reads your code at rest. Its runtime sibling, overloop, is a Claude Code hook that catches the same waste while an agent runs — the tool calls it repeats, the files it re-reads, the oversized output it floods context with. overllm catches the calls you didn't need; overloop catches the ones your agent runs twice. Two halves of the same idea.

Contributing

The most useful thing you can send is a false positive: a real line of code where overllm flags a call it should not. A linter is only worth running if it is right, so one concrete bad flag is worth more than a feature request. CONTRIBUTING.md covers how to report one and how to run the tests.

Past releases and what changed are in CHANGELOG.md.

License

MIT © Adam Danielsson