Project file manager for Claude AI integrations.
Clod (Claude Loader) is a utility for preparing and uploading files to Claude AI's Project Knowledge feature. It tracks file changes, respects .gitignore and .clodignore patterns, and optimizes filenames for Claude's UI.
Key features:
Process all files on first run, only modified files on subsequent runs
Respect .gitignore and .clodignore patterns
Handle binary vs. text files automatically
Use system temporary directories for staging files
Create optimized filenames for Claude's UI
Generate a path manifest for mapping optimized names back to original paths
Color-coded, user-friendly terminal interface
Path-restricted file access to prevent unauthorized operations
Clod is particularly useful for reducing AI development costs while working with Claude. By handling file selection, staging, and tracking efficiently, it can cut API costs by 50% or more. This makes powerful AI tools accessible to students, bootstrappers, and developers on tight budgets, leveling the playing field between the wealthiest and the scrappiest.
Clod implements a capability-based security model to ensure safe AI interactions with the file system, and uses checksum-based file tracking with XXH3 hashes for detecting modified or renamed files. It uses libmagic for robust, content-based file type detection.
Clod - Claude Loader
Clod is a utility for preparing and uploading files to Claude AI's Project Knowledge feature. It tracks file changes using checksums, respects .gitignore
and .clodignore
patterns, optimizes filenames for Claude's UI and provides a filename manifest so Claude can write the files back to their original locations. By efficiently handling file selection and staging, it can significantly reduce AI development costs by 50% or more. Unlike other tools created to solve this problem clod
does not require any unauthorized access to Anthropic products, nor is it affected by changes to Claude's UI. Contributions of Automator code to handle the drag and drops and Project Knowledge deletes on macOS are welcome, as is similar code for other platforms.
A typical workflow for me is to work with Claude App on a new feature for a few iterations, until he gets stuck trying to get a test to pass or something. Then I hand it over to Claude Code and let him solve the final problems with local access and also have him double-check the other Claude’s work. I say “save 50%” but really I probably save more like 90% with this approach.
Claude App is more conversational than Claude Code and he can render web pages and SVGs and so on. Claude Code will often hit a minor roadblock when implementing something and turn around and do exactly what you told him not to do--this is rarely an issue with Claude App as there's more human in the loop. Coding with Claude App isn't just about cost savings for me--I genuinely prefer it for many use cases.
While this project was built with filesystem access in mind I often find myself instead having Claude generate a shell script for me that will apply his changes to the codebase, then I review the script, request changes if needed, download it through the UI and run it. I usually have his script create a tmp directory, create the needed patch files (actual patch files for patch(1) to apply--sometimes Claude just rolls his own patch system if you're not careful) in it, apply the patches, remove the tmp directory if all the patches apply cleanly. Claude usually defaults to sed
for this sort of thing, but it's error prone even for the best frontier AI.
Some people entities, when confronted with a problem, think “I know, I’ll use regular expressions.” Now they have two problems. --Jamie Zawinski
Added April 16, 2025: Claude App has a new 'select all' feature in the Project Knowledge section that now makes the easiest workflow to just delete all Project Knowledge files and replace them with the contents of clod --all
. Note this approach could have an impact on your usage limits, although as of today I have not observed this to be the case.
Developed by Fuzz, Inc - World-class technical leadership and execution
Features
- Track modified files using checksums for accuracy
- Detect renamed files by matching content checksums
- Respect
.gitignore
and.clodignore
patterns - Handle binary vs. text files automatically
- Use system temporary directories for staging files
- Create optimized filenames for Claude's UI
- Generate a path manifest for mapping optimized names back to original paths
- Capability-based security for file operations
- Simple, well-structured monad stack for reliable behavior
Installation
Homebrew (binary is Apple Silicon only for now)
# On macOS
brew tap fuzz/tap
brew install clod
From Hackage
cabal install clod
From Source
git clone https://github.com/fuzz/clod.git
cd clod
cabal install
The clod
program is installed automatically when using cabal install
.
Prerequisites
- Claude Pro, Max, Teams or Enteprise account
- Claude desktop app for filesystem access (currently only macOS and Windows)
- GHC (Glasgow Haskell Compiler) 9.0 or newer
- libmagic (required for file type detection)
Cross-Platform Support: Clod has been tested on macOS, but should work on Linux and Windows. The program outputs the path to the staging directory, making it easy to open with your system's file browser or use with any command that accepts a directory path.
- macOS:
open
- Linux:
xdg-open
,gio
,gnome-open
, orkde-open
- Windows:
explorer.exe
Pull requests for improved cross-platform support are welcome.
Usage
Basic Usage
# Process all files (first run) or modified files since last run
clod
# Process all files regardless of last run (respecting .gitignore and .clodignore)
clod --all
# On macOS, process files and open the staging directory in Finder
open `clod`
Command-Line Options
--all
,-a
: Process all files, not just modified ones--test
,-t
: Run in test mode (no prompts, useful for CI)--staging-dir DIR
,-d DIR
: Specify a directory for test mode (only used with --test)--verbose
,-v
: Enable verbose output--flush
,-f
: Flush stale entries from the checksums database--last
,-l
: Reuse the previous staging directory--help
: Show help information--version
,-V
: Show version information
Opening the Staging Directory
Clod outputs the path to the staging directory, which you can use to open it directly in your file browser:
# On macOS, process files and open the directory in Finder
open `clod`
# For scripts, you can capture the output and open it with your preferred application
STAGING_DIR=$(clod [options])
# Open with the appropriate command for your platform
# macOS
open "$STAGING_DIR"
# Linux
xdg-open "$STAGING_DIR" # or gio, gnome-open, kde-open
# Windows
explorer.exe "$STAGING_DIR"
First Run
On first run, Clod will:
- Create a system temporary directory for staging files
- Create a default
.clodignore
file if one doesn't exist - Prompt you to choose which files to process:
- All files
- Only modified files
- None (just set timestamp)
Integration with Claude AI
First time: Paste the contents of project-instructions.md
into the Project Instructions section
After running Clod:
- Navigate to Project Knowledge in your Claude Project (Pro or Team account required)
- Drag files from the opened staging folder to Project Knowledge
- Include the
_path_manifest.dhall
file which maps optimized names back to original paths - Important: You must manually delete previous versions of these files from Project Knowledge before starting a new conversation to ensure Claude uses the most recent files
- Note that the staging directory is temporary and will be cleaned up on your next run of clod (or system reboot)
Configuration
Environment Variables
You can customize Clod's behavior using these environment variables:
CLOD_DIR
- Override the default.clod
directory nameCLODIGNORE
- Override the default.clodignore
filename
.clodignore
A .clodignore
file in your repository root specifies files or patterns to exclude. If this file doesn't exist, Clod will create a default one for you with common patterns for binary files, build directories, and large files.
Development Utilities
The Clod package includes a testing utility:
magictest
A simple utility to test the libmagic dependency:
cabal run magictest -- /path/to/file
The magictest
tool uses the libmagic library to analyze a file and determine its MIME type and encoding. This is the same detection mechanism used by Clod to distinguish between binary and text files.
Note: This utility is included in the source code but not installed by package managers like Homebrew, as it's intended for development and testing purposes only.
Architecture
Clod uses a clean, pragmatic architecture with a focus on reliability and maintainability:
- Clean Monad Stack: Uses a ReaderT/ExceptT/IO pattern for clear error handling
- Capability-Based Security: Runtime checking of file access permissions based on explicitly allowed directories
- Modular Design: Clear separation of concerns between different subsystems
- Safety First: Designed to prevent accidental access to unauthorized files
The architecture focuses on reliability and maintainability, delivering a system that works effectively with clear error messages.
Project Structure
app/
: Application entry pointsrc/
: Source code modulesClod/Config.hs
: Environment and configuration handlingClod/Core.hs
: Main functionalityClod/FileSystem.hs
: File operations facadeClod/FileSystem/Detection.hs
: File type detectionClod/FileSystem/Operations.hs
: Basic file operationsClod/FileSystem/Processing.hs
: File processing and manifest generationClod/FileSystem/Transformations.hs
: Special file format transformationsClod/FileSystem/Checksums.hs
: Checksum-based file tracking
Clod/IgnorePatterns.hs
: Pattern matchingClod/Output.hs
: User interfaceClod/Types.hs
: Core types and monad stackClod/Effects.hs
: Effect system support
test/
: Test suite.clod/
: Configuration and state (created during execution)
License
This project is licensed under the MIT License - see the LICENSE file for details.