Overview

lmpack packs a project into one file for a language model. It walks the tree, selects files by globs and .gitignore, counts tokens with a real tokenizer, fits the result into a token budget, and prints a report of what was left out and why.

npx lmpack pack --budget 120k --out pack.md

The problem

To show a project to a model, people open a dozen files and paste them into a chat. Sooner or later a node_modules file or a database dump slips in, the context limit is hit, and pieces get thrown out at random. The next day the same selection cannot be repeated, and nobody knows what the model did not see.

lmpack does this in one command, the same way every time, and writes down what it dropped.

What it does

• Fits a token budget. --budget 120k drops files by category priority, largest first, then puts back whatever still fits. The final pack is re-counted, so it never exceeds the budget. See Budget.
• Reports what was left out. Every skipped file gets a line with a reason: budget, size, binary, secret, ignored, lock file, symlink.
• Packs only what changed. --git-diff main selects files changed since the branch point, plus files in the same directories. See Recipes.
• Filters secrets by default. .env, *.pem, id_*, PEM private keys and others never reach the pack unless you pass --allow-secrets. See Secret filter.
• Repeats a selection. Profiles in lmpack.json give the same result today and next month. See Profiles.
• Works offline. No network access, no telemetry, no update checks.

What it is not

Not a code index or a search engine. Not RAG: no embeddings, no state between runs. Not an API client: the pack goes to a file or stdout, and you decide where to send it.

Compared to other tools

Several tools pack a repository for a model. The table covers what each one documents; checked against their READMEs and docs on 2026-09-26. If something here is out of date, please open an issue.

	lmpack	repomix	files-to-prompt	code2prompt
Runtime	Node.js	Node.js	Python	Rust (+ Python SDK)
Output formats	markdown, xml	xml, markdown, json, plain	plain, markdown, Claude XML	markdown, json, xml, Handlebars templates
Token counting	o200k, cl100k, p50k, r50k	o200k, cl100k	no	tiktoken encodings
Token budget	drops files to fit, exit code 2	--token-budget fails when exceeded, output unchanged	no	no
Report of skipped files	every file with a reason	suspicious files	no	no
Git	only changed files (+ neighbors)	adds diff and log sections	no	adds diff and log between branches
Secret filter	on by default: file names, PEM keys and certificates	on by default: Secretlint content scan	no	no
Code compression	no	Tree-sitter --compress	no	no
Remote repositories	no	--remote	no	no
MCP server	no	yes	no	yes

Where the others are stronger:

• repomix is the most complete tool in this space. Its secret check scans file contents with Secretlint and catches API keys, tokens and connection strings inside ordinary files; lmpack only looks at file names and PEM armor. It also compresses code with Tree-sitter, packs remote repositories, splits output, and runs as an MCP server.
• code2prompt has Handlebars templates, a TUI and a token map of the tree.
• files-to-prompt is a small, dependable Python tool; if you do not need token counts, it is enough.

Where lmpack is different:

• The budget is enforced by dropping files in a defined order, and the report says which files went and how many tokens that freed. The others either do not count against a limit or only fail when it is exceeded.
• --git-diff changes which files are packed. The others add a diff section next to the full selection.
• Exit code 2 lets a script notice that the pack is incomplete.

ctxpack is sometimes mentioned alongside these, but it solves a different problem: it indexes repositories, runs semantic search, and answers questions through a connected model provider. lmpack produces a file and never talks to a model.

Next

• Install and run the quick start.
• Look up every flag in the CLI reference.