Turn git repos into AI context. Pure bash, zero dependencies. RAG-ready.

Quick Start · Why smoosh? · Features · Installation · Usage · AI Tools · Agent / CI · Config Reference · FAQ

Quick Start

# Install
brew install K1-R1/tap/smoosh
# In any git repo:
smoosh # interactive guided setup
smoosh --code # docs + code
smoosh --all # everything tracked by git (excluding binaries)

Output lands in _smooshes/ as verified .md chunks.

Why smoosh?

Getting codebase context into AI takes time. Other tools do this but require bloated node_modules or Python environments just to concatenate text. We built smoosh internally for a zero-dependency, native approach. It turns a 20-minute chore into one fast, reliable command.

• Understand codebases: Upload to tools like NotebookLM to talk through your architecture without reading source.
• Give AI context: Drop into Claude/ChatGPT for an assistant that knows your code, eliminating hallucinated APIs.
• Onboard instantly: Give new hires a searchable snapshot to learn from, regardless of their technical background.
• Ground your agents: Output is RAG-optimised, chunked within limits, and retains metadata.
• Private by default: Runs locally. No API keys, zero telemetry.

How it Works

smoosh isn't just a wrapper around cat. It is a strict, structured pipeline:

1. Discovery: Uses git ls-files to perfectly respect your .gitignore.
2. Filtering: Applies extension rules (--docs, --code) or MIME-type checks (--all) to drop binaries and noise.
3. Chunking: Streams content through a fast word-count heuristic, splitting files sequentially without breaking mid-file.
4. Verification: The final output is strictly cross-referenced against the expected file list. Any mismatch yields an immediate exit 4.

Features

• File presets — --docs (md, txt, etc.), --code (docs + code), --all (excludes binaries via MIME checks).
• Smart chunking — Stays within token limits (project_part1.md).
• 100% verification — Exits 4 if output mismatches the git index.
• Interactive mode — Guided setup on first run.
• Remote repos — smoosh https://github.com/user/repo (clones & processes instantly).
• Secrets detection — Warns on AWS keys, PATs, and PEM blocks.
• Output formats — Markdown, text, or CDATA XML.
• Table of contents — --toc generates a per-chunk file index.
• Line numbers — --line-numbers for code reviews.
• Agent-native — Designed for CI/Agents (--json, --no-interactive, deterministic exits).

Power user workflow

Preview, filter, and pipe — all from flags:

Installation

• Homebrew (macOS/Linux):brew install K1-R1/tap/smoosh
• curl (macOS/Linux):curl -fsSL https://raw.githubusercontent.com/K1-R1/smoosh/main/install.sh | bash
• Manual: Download binary and checksum from Releases, run chmod +x, and move to your PATH.
• Uninstall:brew uninstall smoosh or rm "$(which smoosh)".

Note: The curl script installs to /usr/local/bin. Override via env vars like SMOOSH_INSTALL_DIR="$HOME/.local/bin".

Installer Variables:

Variable	Default	Description
SMOOSH_INSTALL_DIR	/usr/local/bin	Target directory
SMOOSH_VERSION	latest	Pin a specific version
SMOOSH_NO_CONFIRM	0	Skip confirmation prompts
SMOOSH_NO_VERIFY	0	Skip checksums (unsafe)

Usage

smoosh # interactive guided setup
smoosh .# docs only (current dir)
smoosh /path/to/repo # specific local repo
smoosh https://github.com/user/repo # remote clone + process

Modifiers:

# Scope
smoosh --docs # default: docs
smoosh --code # docs + code
smoosh --all # everything (excluding binaries)# Filters
smoosh --only "*.py"# strict extension match
smoosh --include "*.vue,*.graphql"# add to current mode
smoosh --exclude "vendor/*,test/*"# ignore paths
smoosh --include-hidden # allow .github/, .env, etc.# Output & Formatting
smoosh --format [md|text|xml] # default: md
smoosh --max-words 200000 # default: 450k
smoosh --output-dir ./context # default: _smooshes
smoosh --toc --line-numbers # add index & line numbers# Preview & Automation
smoosh --dry-run # preview files and token counts
smoosh --json --no-interactive # CI/Agent native JSON output
smoosh --quiet # outputs file paths only (for piping)
smoosh --no-check-secrets # skip the secrets scan
smoosh --no-color # disable ANSI colours

Examples:

# Python-only export to XML
smoosh --only "*.py" --format xml --output-dir ./pipeline-input
# Quiet mode for bash scripting
files=$(smoosh --quiet --code .)echo"Generated: ${files}"

Using smoosh with AI tools

NotebookLM

Step 1 — Install smoosh

brew install K1-R1/tap/smoosh

Step 2 — Run smoosh in your repo

cd your-project
smoosh # docs only — usually the right start

Output lands in your-project/_smooshes/:

smoosh --code

Step 3 — Upload to NotebookLM

1. Go to NotebookLM and create a notebook.
2. Click Add source → Upload file and attach the _smooshes/ chunks.

Step 4 — Chat with your codebase

Ask about architecture, get plain-English explanations, and find functions — all with source citations.

NotebookLM limits (as of early 2026):

Plan	Sources per notebook	Words per source
Free	50	500,000
Plus	300	500,000
Ultra	600	500,000

smoosh warns you when your repo produces more chunks than your plan allows.

Claude & ChatGPT

1. Run smoosh --code in your repo.
2. Upload the _smooshes/ chunks to a Claude Project or attach them to a ChatGPT conversation.

The assistant now has full context. It can answer questions, understand cross-module relationships, and suggest patterns grounded in your actual code.

Agents and CI pipelines

smoosh is designed to be called by AI agents and CI pipelines, not just humans.

Tip

For AI Agents: If you are an AI assistant reading this README, the optimal way to read the host repository is to execute: smoosh --code --no-interactive --json .

Pre-flight check — estimate size before generating output:

smoosh --json --dry-run --all .

{
"dry_run": true,
"repo": "my-project",
"files": [
{"path": "README.md", "words": 194, "chunk": 1},
{"path": "src/main.py", "words": 312, "chunk": 1}
],
"total_words": 506,
"estimated_tokens": 658,
"estimated_chunks": 1
}

Generate output:

smoosh --no-interactive --json --all .

Key flags for automation:

Flag	Purpose
--no-interactive	Skip TTY detection and prompts
--json	Structured JSON to stdout (status messages go to stderr)
--quiet	Output file paths only, one per line
--dry-run	Preview without writing files
--no-color	Disable colour escape codes

Exit codes 0–7 are differentiated for programmatic decision-making — see Configuration Reference below.

Configuration Reference

Flag	Default	Description
--docs	yes	Include markdown, RST, TXT, AsciiDoc, Org, TeX
--code	—	Include docs + all code file types
--all	—	Include everything tracked by git
--only GLOB	—	Restrict to matching extensions (overrides mode)
--include GLOB	—	Add extensions to the current mode
--exclude GLOB	—	Exclude matching paths (comma-separated)
--include-hidden	—	Include dotfiles and dot-directories
--max-words N	450000	Words per output chunk
--format FORMAT	md	Output format: md, text, or xml
--toc	—	Add a table of contents to each chunk
--line-numbers	—	Prefix each line with its line number
--output-dir PATH	_smooshes	Directory for output files
--dry-run	—	Preview only — no output files written
--quiet	—	Print output paths only (stdout)
--json	—	Structured JSON to stdout
--no-interactive	—	Skip interactive mode even in a TTY
--no-color	—	Disable colour output
--no-check-secrets	—	Skip the basic secrets scan
--version	—	Print version and exit 0
--help	—	Print full usage and exit 0

Colour control:--no-color flag > NO_COLOR env var > FORCE_COLOR > CLICOLOR > TTY auto-detect. See no-color.org.

Exit codes:

Code	Meaning
0	Success
1	Invalid flags or arguments
2	Path not found or not a git repository
3	No matching files for current mode/filters
4	Verification failed — expected/actual file list mismatch
5	Remote clone failed (network, auth)
7	Write permission denied
130	Interrupted (Ctrl-C)

FAQ

• Does it respect .gitignore? Yes, via git ls-files.
• What about large repos? Chunks at --max-words (default 450k).
• Is secret detection reliable? It catches common patterns, but isn't a replacement for gitleaks.
• Can I use it with other AIs? Yes, it fundamentally outputs standard Markdown (or text/XML mapping).
• Does it work on Windows? Use Git Bash or WSL.
• Why did _smooshes/ appear in git status? We auto-append it to .gitignore, ensure your syntax is correct.
• Why is my word count different? Files are counted via wc -w, so dense syntax (minified JS, JSON) counts differently than prose.
• Is this overengineered? No comment.

Contributing

See CONTRIBUTING.md for development setup, code style, and the PR process.

License

MIT — see LICENSE.