Jinni: Bring Your Project Into Context

Jinni is a tool to efficiently provide Large Language Models the context of your projects. It gives a consolidated view of relevant project files, overcoming the limitations and inefficiencies of reading files one by one. Each file's content is preceded by a simple header indicating its path:

```path=src/app.py
print("hello")

The philosophy behind this tool is that LLM context windows are large, models are smart, and directly seeing your project best equips the model to help with anything you throw at it.

There is an MCP (Model Context Protocol) server for integration with AI tools and a command-line utility (CLI) for manual use that copies project context to the clipboard ready to paste wherever you need it.

These tools are opinionated about what counts as relevant project context to best work out of the box in most use cases, automatically excluding:

* Binary files
* Dotfiles and hidden directories
* Common naming conventions for logs, build directories, tempfiles, etc

Inclusions/exclusions are customizable with complete granularity if required using .contextfiles – this works like .gitignore except defining inclusions. .gitignore files themselves are also respected automatically, but any rules in .contextfiles take priority.

The MCP server can provide as much or as little of the project as desired. By default the scope is the whole project, but the model can ask for specific modules / matching patterns / etc.

MCP Quickstart

MCP server config file for Cursor / Roo / Claude Desktop / client of choice:

{
    "mcpServers": {
        "jinni": {
            "command": "uvx",
            "args": ["jinni-server"]
        }
    }
}

You can optionally constrain the server to only read within a tree for security in case your LLM goes rogue: add "--root", "/absolute/path/" to the args list.

Install uv if it is not on your system: https://docs.astral.sh/uv/getting-started/installation/

Reload your IDE and you can now ask the agent to read in context.

If you want to restrict this to particular modules / paths just ask - e.g. "Read context for tests".

In action with Cursor:

Note For Cursor Users

Cursor can silently drop context that is larger than the allowed maximum, so if you have a sizable project and the agent acts like the tool call never happened, try reducing what you are bringing in ("read context for xyz")

Components

1. jinni MCP Server:

   • Integrates with MCP clients like Cursor, Cline, Roo, Claude Desktop, etc.
   • Exposes a read_context tool that returns a concatenated string of relevant file contents from a specified project directory.

2. jinni CLI:

   • A command-line tool for manually generating the project context dump.
   • Useful for feeding context to LLMs via copy-paste or file input. Or pipe the output wherever you need it.

Features

• Efficient Context Gathering: Reads and concatenates relevant project files in one operation.
• Intelligent Filtering (Gitignore-Style Inclusion):
  • Uses a system based on .gitignore syntax (pathspec library's gitwildmatch).
  • Automatically loads .gitignore files from the project root downward. These exclusions can be overridden by rules in .contextfiles.
  • Supports hierarchical configuration using .contextfiles placed within your project directories. Rules are applied dynamically based on the file/directory being processed.
  • Matching Behavior: Patterns match against the path relative to the target directory being processed. Output paths remain relative to the original project root.
  • Rule Root Behavior: Each target has its own rule root:
    • Targets within the project root (or CWD) use the project root/CWD as their rule root
    • External targets use themselves as their rule root, ensuring self-contained rule sets
  • Overrides: Supports --overrides (CLI) or rules (MCP) to use a specific set of rules exclusively. When overrides are active, both built-in default rules and any .contextfiles are ignored. Path matching for overrides is still relative to the target directory.
  • Explicit Target Inclusion: Files explicitly provided as targets are always included (bypassing rule checks, but not binary/size checks).
• Customizable Configuration (.contextfiles / Overrides):
  • Define precisely which files/directories to include or exclude using .gitignore-style patterns applied to the relative path.
  • Patterns starting with ! negate the match (an exclusion pattern). (See Configuration section below).
• Large Context Handling: Aborts with a DetailedContextSizeError if the total size of included files exceeds a configurable limit (default: 100MB). The error message includes a list of the 10 largest files contributing to the size, helping you identify candidates for exclusion. See the Troubleshooting section for guidance on managing context size.
• Metadata Headers: Output includes a path header for each included file (e.g., ````path=src/app.py). This can be disabled with list_only`.
• Encoding Handling: Attempts multiple common text encodings (UTF-8, Latin-1, etc.).
• List Only Mode: Option to only list the relative paths of files that would be included, without their content.

Platform-specific notes

Windows + WSL

Jinni v0.1.7+ auto-converts WSL paths.

Provide either of these as project_root (CLI --root or MCP argument):

/home/user/project
vscode-remote://wsl+Ubuntu-22.04/home/user/project

No wrappers, mounts, or extra flags required—Jinni resolves the UNC path (\\wsl$\...) on Windows automatically.

UNC Path Format: Jinni always uses \\wsl$\<distro>\... for maximum compatibility with all Windows versions supporting WSL. Distro Name Handling: Spaces and most special characters are allowed in the distro name. Only truly illegal UNC characters are replaced with _. Caching: WSL path lookups and conversions are cached for performance. If you install WSL while Jinni is running, restart Jinni to pick up the new wslpath. Opt-out: Set the environment variable JINNI_NO_WSL_TRANSLATE=1 to disable all WSL path translation logic.

Only wsl+<distro> URIs and absolute POSIX paths (starting with /) are translated; for SSH or container remotes, run Jinni inside that environment.

Examples

• Dump context of my_project/ to the console:

  Copy

  jinni ./my_project/ # Process a single directory
  jinni ./src ./docs/README.md # Process multiple targets
  jinni # Process current directory (.)

• List files that would be included in my_project/ without content:

  Copy

  jinni -l ./my_project/
  jinni --list-only ./src ./docs/README.md

• Dump context of my_project/ to a file named context_dump.txt:

  Copy

  jinni -o context_dump.txt ./my_project/

• Use override rules from custom.rules instead of .contextfiles:

  Copy

  jinni --overrides custom.rules ./my_project/

• Show debug information:

  Copy

  jinni --debug-explain ./src

• Dump context (output is automatically copied to clipboard by default):

  Copy

  jinni ./my_project/

• Dump context but do not copy to clipboard:

  Copy

  jinni --no-copy ./my_project/

Development

• Design Details: DESIGN.md

• Running Server Locally: During development (after installing with uv pip install -e . or similar), you can run the server module directly:

  Copy

  python -m jinni.server [OPTIONS]

  Example MCP client configuration for local development:

  Copy

  {
    "mcpServers": {
      "jinni": {
        // Adjust python path if needed, or ensure the correct environment is active
        "command": "python -m jinni.server"
        // Optionally constrain the server to only read within a tree (recommended for security):
        // "command": "python -m jinni.server --root /absolute/path/to/repo"
      }
    }
  }

Usage

MCP Server (read_context tool)

1. Setup: Configure your MCP client (e.g., Claude Desktop's claude_desktop_config.json) to run the jinni server via uvx.
2. Invocation: When interacting with your LLM via the MCP client, the model can invoke the read_context tool.
   • project_root (string, required): The absolute path to the project root directory. Rule discovery and output paths are relative to this root.
   • targets (JSON array of strings, required): Specifies a mandatory list of file(s)/directory/ies within project_root to process. Must be a JSON array of string paths (e.g., ["path/to/file1", "path/to/dir2"]). Paths can be absolute or relative to CWD. All target paths must resolve to locations inside project_root. If an empty list [] is provided, the entire project_root is processed.
   • rules (JSON array of strings, required): A mandatory list of inline filtering rules (using .gitignore-style syntax, e.g., ["src/**/*.py", "!*.tmp"]). Provide an empty list [] if no specific rules are needed (this will use built-in defaults). If non-empty, these rules are used exclusively, ignoring built-in defaults and .contextfiles.
   • list_only (boolean, optional): If true, returns only the list of relative file paths instead of content.
   • size_limit_mb (integer, optional): Override the context size limit in MB.
   • debug_explain (boolean, optional): Enable debug logging on the server.
   • exclusions (object, optional): Exclusion configuration with three optional fields:
     • global (array of strings): Keywords to exclude globally (e.g., ["tests", "deprecated"])
     • scoped (object): Map of paths to keyword arrays for scoped exclusions (e.g., {"src/legacy": ["old", "deprecated"]})
     • patterns (array of strings): File patterns to exclude (e.g., ["*.test.js", "*_old.*"])
   3. Output: The tool returns a single string containing the concatenated content (with headers) or the file list. Paths in headers/lists are relative to the provided project_root. In case of a context size error, it returns a DetailedContextSizeError with details about the largest files.

MCP Server (usage tool)

• Invocation: The model can invoke the usage tool (no arguments needed).
• Output: Returns the content of the README.md file as a string.

(Detailed server setup instructions will vary depending on your MCP client. Generally, you need to configure the client to execute the Jinni server.)

Running the Server:

• Recommended Method: Use uvx to run the server entry point directly (requires the jinni package to be published on PyPI or findable by uvx):
  Copy

  uvx jinni-server [OPTIONS]

  Example MCP client configuration (e.g., claude_desktop_config.json):
  Copy

  {
    "mcpServers": {
      "jinni": {
        "command": "uvx",
        "args": ["jinni-server"]
      }
    }
  }

You can optionally constrain the server to only read within a tree for security in case your LLM goes rogue: add "--root", "/absolute/path/" to the args list.

See your specific MCP client's documentation for precise setup steps. Ensure uv is installed

Command-Line Utility (jinni CLI)

jinni [OPTIONS] [<PATH...>]

• <PATH...> (optional): One or more paths to the project directories or files to analyze. Defaults to the current directory (.) if none are provided.
• -r <DIR> / --root <DIR> (optional): Specify the project root directory. If provided, rule discovery starts here, and output paths are relative to this directory. If omitted, the root is inferred from the common ancestor of the <PATH...> arguments (or CWD if only '.' is processed).
• --output <FILE> / -o <FILE> (optional): Write the output to <FILE> instead of printing to standard output.
• --list-only / -l (optional): Only list the relative paths of files that would be included.
• --overrides <FILE> (optional): Add rules from <FILE> as high-priority rules in addition to .contextfiles and .gitignore.
• --size-limit-mb <MB> / -s <MB> (optional): Override the maximum context size in MB.
• --debug-explain (optional): Print detailed inclusion/exclusion reasons to stderr and jinni_debug.log.
• --root <DIR> / -r <DIR> (optional): See above.
• --no-copy (optional): Prevent automatically copying the output content to the system clipboard when printing to standard output (the default is to copy).
• --not <keyword> (optional, repeatable): Exclude modules/directories matching keyword (e.g., --not tests --not vendor). Can be used multiple times.
• --not-in <path:keywords> (optional, repeatable): Exclude specific keywords within a path (e.g., --not-in src/legacy:old,deprecated). Can be used multiple times.
• --not-files <pattern> (optional, repeatable): Exclude files matching pattern (e.g., --not-files '*.test.js' --not-files '*_old.*'). Can be used multiple times.
• --keep-only <modules> (optional): Keep only specified modules/directories, exclude everything else (comma-separated, e.g., --keep-only src,lib,docs).

Exclusion Examples

CLI Examples:

# Exclude all test directories
jinni --not tests

# Exclude multiple keywords
jinni --not tests --not vendor --not deprecated

# Exclude old code only in specific paths
jinni --not-in src/legacy:old,deprecated --not-in lib/v1:legacy

# Exclude specific file patterns
jinni --not-files "*.test.js" --not-files "*_old.*"

# Keep only src and docs, exclude everything else
jinni --keep-only src,docs

# Combine different exclusion types
jinni --not tests --not-in src/experimental:wip --not-files "*.bak"

Note: Exclusion commands (--not* flags) work in addition to existing .gitignore and .contextfiles rules. They further filter down what would otherwise be included.

MCP Examples:

{
  "project_root": "/path/to/project",
  "targets": [],
  "rules": [],
  "exclusions": {
    "global": ["tests", "vendor"],
    "scoped": {
      "src/legacy": ["old", "deprecated"],
      "lib/experimental": ["wip", "unstable"]
    },
    "patterns": ["*.test.js", "*_backup.*"]
  }
}

Installation

You can install Jinni using pip or uv:

Using pip:

pip install jinni

Using uv:

uv pip install jinni

This will make the jinni CLI command available in your environment. See the "Running the Server" section above for how to start the MCP server depending on your installation method.

Configuration (.contextfiles & Overrides)

Jinni uses .contextfiles (or an override file) to determine which files and directories to include or exclude, based on .gitignore-style patterns.

• Core Principle: Rules are applied dynamically during traversal, relative to the current target directory being processed.
• Location (.contextfiles): Place .contextfiles in any directory. Rule discovery starts from the rule root (project root for internal targets, target itself for external targets) and proceeds downward to the current directory being processed.
• Format: Plain text, UTF-8 encoded, one pattern per line.
• Syntax: Uses standard .gitignore pattern syntax (specifically pathspec's gitwildmatch implementation).
  • Comments: Lines starting with # are ignored.
  • Inclusion Patterns: Specify files/directories to include (e.g., src/**/*.py, *.md, /config.yaml).
  • Exclusion Patterns: Lines starting with ! indicate that a matching file should be excluded (negates the pattern).
  • Anchoring: A leading / anchors the pattern to the directory containing the .contextfiles.
  • Directory Matching: A trailing / matches directories only.
  • Wildcards: *, **, ? work as in .gitignore.
• Rule Application Logic:
  1. Determine Target: Jinni identifies the target directory (either explicitly provided or the project root).
  2. Override Check: If --overrides (CLI) or rules (MCP) are provided, these rules are used exclusively. All .contextfiles and built-in defaults are ignored. Path matching is relative to the target directory.
  3. Dynamic Context Rules (No Overrides): When processing a file or subdirectory:
     • Jinni finds all .gitignore and .contextfiles starting from the rule root down to the current item's directory.
     • Rules are combined in order: built-in defaults, .gitignore rules, .contextfiles rules (which take precedence).
     • It compiles these combined rules into a specification (PathSpec).
     • It matches the current file/subdirectory path, calculated relative to the target directory, against this specification.
  4. Matching: The last pattern in the combined rule set that matches the item's relative path determines its fate. ! negates the match. If no user-defined pattern matches, the item is included unless it matches a built-in default exclusion (like !.*).
  5. Target Handling: Explicitly targeted files bypass rule checks. Output paths always remain relative to the original project_root.

Examples (.contextfiles)

Example 1: Include Python Source and Root Config

Located at my_project/.contextfiles:

# Include all Python files in the src directory and subdirectories
src/**/*.py

# Include the main config file at the root of the project
/config.json

# Include all markdown files anywhere
*.md

# Exclude any test data directories found anywhere
!**/test_data/

Example 2: Overriding in a Subdirectory

Located at my_project/src/.contextfiles:

# In addition to rules inherited from parent .contextfiles...

# Include specific utility scripts in this directory
utils/*.sh

# Exclude a specific generated file within src, even if *.py is included elsewhere
!generated_parser.py

Troubleshooting

Context Size Errors (DetailedContextSizeError)

If you encounter an error indicating the context size limit was exceeded, Jinni will provide a list of the 10 largest files it attempted to include. This helps you identify potential candidates for exclusion.

To resolve this:

1. Review the Largest Files: Check the list provided in the error message. Are there large files (e.g., data files, logs, build artifacts, media) that shouldn't be part of the LLM's context?
2. Configure Exclusions: Use .contextfiles or the --overrides / rules options to exclude unnecessary files or directories.
   • Example (.contextfiles): To exclude all .log files and a specific large data directory:
     Copy

     # Exclude all log files
     !*.log
     
     # Exclude a large data directory
     !large_data_files/

   • Refer to the Configuration section above for detailed syntax and usage.
3. Increase the Limit (Use with Caution): If all included files are genuinely necessary, you can increase the size limit using --size-limit-mb (CLI) or size_limit_mb (MCP). Be mindful of LLM context window limits and processing costs.
4. Use jinni usage / usage: If you need to refer back to these instructions or the configuration details while troubleshooting, use the jinni usage command or the usage MCP tool.