llmignore

Like .gitignore, but it tells AI tools what not to read.

Secrets, dependencies, build output and binaries should never reach your AI assistant. .llmignore keeps them out — one file, every tool.

---

$ llmignore scan

⚠ 2 exposed secrets reachable by AI tools:

  .env     Environment file - commonly holds API keys and DB passwords
  id_rsa   SSH private key

Add them to .llmignore (or run `llmignore init`).

Your AI assistant reads your whole repo to "understand the codebase." That includes your .env, your private keys, your 200 MB of node_modules, and your minified bundles — all shipped to a model, all eating your context window. .llmignore fixes that.

Why

• Stop leaking secrets. .env, *.pem, id_rsa, .npmrc — gone from AI context by default.
• Stop wasting context (and money). Skip node_modules/, dist/, lockfiles, binaries. Smaller, sharper context.
• One file, every tool. Write .llmignore once; sync mirrors it to the ignore files 9 AI tools already read - Cursor, Windsurf, Gemini, Aider, Continue, Cline, Roo Code, JetBrains AI Assistant.
• Native speed. Built on ripgrep's engine. Scans tens of thousands of files in milliseconds, single static binary, ~2 MB, zero runtime.

Install

# npm (no install, just run)
npx llmignore-cli init

# npm (global)
npm install -g llmignore-cli

# Cargo
cargo install llmignore-cli

# Homebrew (macOS / Linux)
brew install horiastanxd/tap/llmignore

# Scoop (Windows)
scoop install https://raw.githubusercontent.com/horiastanxd/llmignore/main/scoop/llmignore.json

# curl
curl -fsSL https://raw.githubusercontent.com/horiastanxd/llmignore/main/install.sh | sh

The package is llmignore-cli; the installed command is llmignore.

Quick start

llmignore init     # write a .llmignore with strong defaults
llmignore scan     # any secrets currently exposed to AI? (exit 1 if yes)
llmignore sync     # mirror it into .cursorignore, .aiexclude, ...
llmignore          # summary: what AI sees vs what it skips

$ llmignore

llmignore  ~/code/my-app

  Included       206 files   ~753.7k tokens
  Ignored      6,386 files
  Rules      built-in defaults + .gitignore

  ⚠ 1 exposed secret (run `llmignore scan`)

That's 206 files going to your AI instead of 6,592. Less noise, no secrets, lower cost.

The .llmignore file

Exactly the same syntax as .gitignore — globs, ! negation, per-directory files, the lot:

# .llmignore
.env
.env.*
!.env.example
*.pem
id_rsa
node_modules/
dist/
*.min.js
*.png
package-lock.json

If a repo has no .llmignore, llmignore falls back to a comprehensive built-in default ruleset (secrets, deps, build output, caches, lockfiles, binaries, media, logs). Run llmignore init to drop that ruleset into a file you can edit. .gitignore is honored too, so your existing ignores already apply.

Commands

Command	What it does
llmignore init	Write a .llmignore with sensible defaults (--force to overwrite).
llmignore scan	List secret files currently reachable by AI. Exits 1 if any — drop into CI.
llmignore sync	Generate the ignore files AI tools read: .cursorignore, .codeiumignore, .aiexclude, .geminiignore, .aiderignore, .continueignore, .clineignore, .rooignore, .aiignore.
llmignore completions <shell>	Print a shell completion script (bash/zsh/fish/...).
llmignore list	Print files an AI would read. --ignored for the inverse, --json, -0 for xargs -0.
llmignore check <file>	Will an AI read this file? Exit 0 = yes, 1 = ignored.
llmignore stats	Included vs ignored counts and an estimated token total.

Global flags: -C <dir> (run elsewhere), --no-gitignore, --no-defaults. --json is available on scan, list, and stats for scripting.

Use it today

No assistant reads .llmignore natively yet — so sync bridges the gap by writing the files the tools already respect:

llmignore init && llmignore sync

synced .cursorignore    Cursor
synced .codeiumignore   Windsurf / Codeium
synced .aiexclude       Gemini Code Assist
synced .geminiignore    Gemini CLI
synced .aiderignore     Aider
synced .continueignore  Continue
synced .clineignore     Cline
synced .rooignore       Roo Code
synced .aiignore        JetBrains AI Assistant

Edit .llmignore, re-run llmignore sync, done. Generated files carry a header so re-syncs update cleanly and never clobber a file you wrote by hand (use --force to override).

Guard your repo in CI

As a GitHub Action - fails the build if a secret is reachable by AI tools:

# .github/workflows/llmignore.yml
name: llmignore
on: [push, pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: horiastanxd/llmignore@v0.1

Or in any pipeline: npx llmignore-cli scan.

As a pre-commit hook

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/horiastanxd/llmignore
    rev: v0.1.0
    hooks:
      - id: llmignore-scan

Feed only safe files to any AI CLI

llmignore list -0 | xargs -0 cat | your-ai-cli

The standard

.llmignore is meant to be a simple, shared convention: a gitignore-syntax file that tells AI tools what not to read. If you build an AI coding tool, please read it. The format is just gitignore — there is nothing new to learn. Issues and proposals welcome.

Roadmap

• llmignore bundle — concatenate the included files into one AI-ready document
• Pre-commit hook for scan
• Editor/MCP integrations that read .llmignore directly

Contributing

Built in Rust on the ignore crate. cargo test runs the full suite. PRs welcome — especially new sync targets and additions to the default ruleset.

License

MIT © horiastanxd