code-explore-by-sql

Local stdio MCP server for fast source code navigation using SQLite FTS5 (trigram tokenizer) + bracket skeleton indexing.

• Full-text search: FTS5 with trigram tokenizer for code-symbol-precise search (GetGBuffer, FMaterial, UE_LOG)
• Symbol lookup: read code by qualified name with fuzzy matching (Jump → ACharacter::Jump)
• Bracket skeleton index: lightweight structural indexing via FSM brace matching (no AST parser needed)
• 12 language support: C, C++, C#, Go, HLSL, GLSL, Java, JavaScript, Kotlin, Python, Rust, Swift
• Multi-database: query multiple codebases simultaneously via CODE_SOURCE_DBS
• Token-efficient responses: compact snippets (~2,600 tokens/20 results, 95% reduction vs full file reads)

# Run the MCP server directly (no clone needed)
uvx code-explore-by-sql

# Or install persistently
pip install code-explore-by-sql

# Build index for your codebase
uvx code-explore-by-sql-build-db /path/to/source /path/to/output.db

# Smoke test with limited files
uvx code-explore-by-sql-build-db /path/to/source /path/to/output.db --limit 1000

Performance: ~84,700 files indexed in ~3.3 minutes on a 2-core machine.

if use with AI Agent, copy current directory .claude/code-source-sql-search skill to AI Agent skills directory and modify mcp settings.

Claude Code (.claude/mcp.json):

{
  "mcpServers": {
    "code-source-sql": {
      "command": "uvx",
      "args": ["code-explore-by-sql"],
      "env": {
        "CODE_SOURCE_DB": "/path/to/your/code.db",
        "CODE_SOURCE_DBS": "/path/to/your/code.db:/path/to/another.db"
      }
    }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "code-source-sql": {
      "type": "stdio",
      "command": "uvx",
      "args": ["code-explore-by-sql"],
      "env": {
        "CODE_SOURCE_DB": "/path/to/your/code.db"
      }
    }
  }
}

OpenAI Codex (~/.codex/config.toml):

[mcp_servers.code-source-sql]
command = "uvx"
args = ["code-explore-by-sql"]

[mcp_servers.code-source-sql.env]
CODE_SOURCE_DB = "/path/to/your/code.db"

Hermes Agent (~/.hermes/config.yaml):

mcp_servers:
  code-source-sql:
    command: uvx
    args:
      - code-explore-by-sql
    env:
      CODE_SOURCE_DB: /path/to/your/code.db

Each tool accepts an optional db parameter to select a database by alias. Aliases are derived from database filenames (unreal.db → "unreal"). Use list_databases to discover available aliases.

Simple mode (keyword):

keyword="GetGBuffer"
keyword="FMaterial Render"

Advanced mode (raw_query) — full FTS5 boolean:

raw_query='"GetGBuffer" AND "Emissive"'
raw_query='"Material" NOT "hlsl"'
raw_query='(file_path : "BasePass") AND "roughness"'
raw_query='(module_name : "Renderer") AND "VirtualTexture"'

1. search_fts_tool(keyword) → file candidates + block QNs
2. search_fts_tool(raw_query, file_path filter) → precise block in target file
3. read_symbol(block QN) or read_file_range(file, line) → full code

┌──────────────────────────────────────────────────────────────┐
│                    MCP Server (FastMCP)                       │
├──────────┬──────────┬──────────┬──────────┬──────────────────┤
│ search   │ read     │ read     │ get_dir  │ list             │
│ fts_tool │ _symbol  │ _file    │ _struct  │ _databases       │
│          │          │ _range   │          │                  │
├──────────┴──────────┴──────────┴──────────┴──────────────────┤
│                     Query Pipeline                            │
│  FTS5 trigram → Symbol match → Edge extraction               │
├──────────────────────────────────────────────────────────────┤
│                    SQLite Database                            │
│  file_content + FTS5 │ symbol_index │ strict_edges           │
└──────────────────────────────────────────────────────────────┘

A 6-state finite state machine (CODE, LINE_COMMENT, BLOCK_COMMENT, STRING, CHAR_LITERAL, RAW_STRING) scans source code tracking brace pairs while correctly ignoring braces in comments and strings.

Top-level blocks are classified by a symbol analyzer producing block_type (namespace/class/enum/function/macro) and block_name (qualified name).

Databases are registered via environment variables at server startup:

• CODE_SOURCE_DB — primary database (default when db is omitted)
• CODE_SOURCE_DBS — colon-separated list of additional databases

Aliases are auto-derived from filename stems. Connections are cached with health checks.

# Clone and setup
git clone https://github.com/didi514354875/code-explore-by-sql.git
cd code-explore-by-sql
uv sync --dev

# Run tests
uv run pytest
uv run ruff check .

MIT

Search inspiration from https://github.com/IOchair/SQL-ManyThing