sari 0.1.8

Sari - Local Search MCP Server

Sari (사리)

Sari is a high-performance Local Code Search Agent implementing the Model Context Protocol (MCP). It empowers AI assistants (like Claude, Cursor, Codex) to efficiently navigate, understand, and search large codebases without sending code to external servers.

Key Features:

• ⚡ Fast Indexing: SQLite FTS5 + AST-based symbol extraction.
• 🔍 Smart Search: Hybrid ranking (Keyword + Symbol structure).
• 🧠 Code Intelligence: Call graphs, snippets management, and domain context archiving.
• 🔒 Local & Secure: All data remains on your machine. No external API dependency.

---

🚀 Installation & Setup

Sari supports automatic installation via MCP configuration (Recommended) or manual installation via pip.

Option 1: Automatic Installation (Recommended)

Add the following configuration to your MCP client (Cursor, Claude Desktop, etc.). Sari will be automatically installed (via pip) and updated upon launch.

🍎 macOS / Linux

Cursor / Claude Desktop Config:

{
  "mcpServers": {
    "sari": {
      "command": "bash",
      "args": [
        "-lc",
        "export PATH=$PATH:/usr/local/bin:/opt/homebrew/bin:$HOME/.local/bin && (curl -fsSL https://raw.githubusercontent.com/BaeCheolHan/sari/main/install.py | python3 - -y || true) && exec ~/.local/share/sari/bootstrap.sh --transport stdio"
      ],
      "env": {
        "DECKARD_WORKSPACE_ROOT": "/path/to/your/project",
        "DECKARD_RESPONSE_COMPACT": "1"
      }
    }
  }
}

🪟 Windows (PowerShell)

Cursor / Claude Desktop Config:

{
  "mcpServers": {
    "sari": {
      "command": "powershell",
      "args": [
        "-NoProfile",
        "-ExecutionPolicy", "Bypass",
        "-Command",
        "irm https://raw.githubusercontent.com/BaeCheolHan/sari/main/install.py | python - -y; & $env:LOCALAPPDATA\sari\bootstrap.bat --transport stdio"
      ],
      "env": {
        "DECKARD_WORKSPACE_ROOT": "C:\\path\\to\\your\\project",
        "DECKARD_RESPONSE_COMPACT": "1"
      }
    }
  }
}

---

Option 2: Manual Installation (Pip)

If you prefer to manage the package manually:

# Install from PyPI
pip install sari

# Run MCP Server
python3 -m sari --transport stdio

---

⚙️ Configuration Reference

How to Configure

You can customize Sari's behavior by setting environment variables.

• MCP Config: Add them to the env dictionary in your config.json or config.toml.
• CLI: Prefix the command, e.g., DECKARD_ENGINE_MODE=sqlite sari status.

"env": {
  "DECKARD_WORKSPACE_ROOT": "/path/to/project",
  "DECKARD_ENGINE_TOKENIZER": "cjk"
}

1. Core & System

Essential settings for basic operation.

Variable	Description	Default
DECKARD_WORKSPACE_ROOT	(Required) Absolute path to the project root. Auto-detected if omitted, but recommended to set explicitly.	Auto-detect
DECKARD_DB_PATH	Custom path for the SQLite database file.	~/.local/share/sari/data/<hash>/index.db
DECKARD_CONFIG	Path to a specific config file to load.	~/.config/sari/config.json
DECKARD_RESPONSE_COMPACT	Minify JSON responses (pack format) to save LLM tokens. Set 0 for pretty-print debugging.	1 (Enabled)
DECKARD_FORMAT	Output format for CLI tools. pack (text-based) or json.	pack

2. Search Engine

Settings to tune search quality and backend behavior.

Variable	Description	Default
DECKARD_ENGINE_MODE	Search backend. embedded uses Tantivy (faster, smart ranking), sqlite uses FTS5 (slower, fallback).	embedded
DECKARD_ENGINE_TOKENIZER	Tokenizer strategy. auto (detects), cjk (optimized for KR/CN/JP), latin (standard).	auto
DECKARD_ENGINE_AUTO_INSTALL	Automatically install engine binaries (Tantivy) if missing.	1 (Enabled)
DECKARD_ENGINE_SUGGEST_FILES	File count threshold to suggest upgrading to Tantivy engine in status checks.	10000
DECKARD_LINDERA_DICT_PATH	Path to custom Lindera dictionary for CJK tokenization (Advanced).	-

3. Indexing & Performance

Fine-tune resource usage and concurrency.

Variable	Description	Default
DECKARD_COALESCE_SHARDS	Number of lock shards for indexing concurrency. Increase for massive repos with frequent changes.	16
DECKARD_PARSE_TIMEOUT_SECONDS	Timeout per file parsing in seconds. Set 0 to disable timeout. Prevents parser hangs.	0
DECKARD_PARSE_TIMEOUT_WORKERS	Worker threads for parsing with timeout.	2
DECKARD_MAX_PARSE_BYTES	Max file size to attempt parsing (bytes). Larger files are skipped or sampled.	16MB
DECKARD_MAX_AST_BYTES	Max file size to attempt AST extraction (bytes).	8MB
DECKARD_GIT_CHECKOUT_DEBOUNCE	Seconds to wait after git checkout before starting bulk indexing.	3.0
DECKARD_FOLLOW_SYMLINKS	Follow symbolic links during file scanning. Caution: May cause infinite loops if circular links exist.	0 (Disabled)
DECKARD_READ_MAX_BYTES	Max bytes returned by read_file tool. Prevents context overflow.	1MB

4. Network & Security

Connectivity settings for the daemon.

Variable	Description	Default
DECKARD_DAEMON_HOST	Host address for the background daemon.	127.0.0.1
DECKARD_DAEMON_PORT	TCP port for the daemon.	47779
DECKARD_HTTP_API_PORT	Port for the HTTP API server (optional).	47777
DECKARD_ALLOW_NON_LOOPBACK	Allow connections from non-localhost IPs. Security Risk: Only enable in trusted networks.	0 (Disabled)

5. Advanced / Debug

Developer options for debugging and plugin extension.

Variable	Description	Default
DECKARD_CALLGRAPH_PLUGIN	Python module path for custom static analysis plugin.	-
DECKARD_DRYRUN_LINT	Enable linter checks (ruff/eslint) in dry_run_diff tool.	0
DECKARD_DLQ_POLL_SECONDS	Interval to retry failed indexing tasks (Dead Letter Queue).	60
DECKARD_LOG_LEVEL	Logging verbosity (DEBUG, INFO, WARNING, ERROR).	INFO

---

🛠️ Usage (MCP Tools)

Once connected, your AI assistant can use these tools:

Core Tools

• search: Search for code or documentation using keywords or regex.
• read_file: Read file content (optimized for large files).
• list_files: List files in the repository.
• search_symbols: Find classes, functions, or methods by name.
• read_symbol: Read only the definition of a specific symbol (saves context).

Intelligence Tools

• call_graph: Analyze function call relationships (upstream/downstream).
• save_snippet / get_snippet: Save and retrieve important code blocks with tags.
• archive_context / get_context: Store domain knowledge and design decisions.
• grep_and_read: Search and read top N files in one go (Composite tool).

---

🩺 Troubleshooting

Check Status

You can check the daemon status and indexing progress:

# If installed automatically:
~/.local/share/sari/bootstrap.sh status

# If installed via pip:
sari status

Run Doctor

Diagnose issues with your environment or installation:

# If installed automatically:
~/.local/share/sari/bootstrap.sh doctor --auto-fix

# If installed via pip:
sari doctor --auto-fix

Uninstall

To remove Sari and all indexed data:

# macOS/Linux
curl -fsSL https://raw.githubusercontent.com/BaeCheolHan/sari/main/install.py | python3 - --uninstall

# Windows
irm https://raw.githubusercontent.com/BaeCheolHan/sari/main/install.py | python - --uninstall

---

📜 License

Apache License 2.0