Universal code-indexer MCP server for AI coding agents.
Claude · Cursor · Codex CLI · Gemini CLI · GitHub Copilot · Windsurf · Kilo Code · Zed · OpenCode · Any agent that speaks MCP.
Pre-index your codebase once. Let AI agents find any symbol in ~200 tokens instead of reading whole files at ~7,500 tokens.
That is a 97% reduction — per lookup, every lookup.
The part no other code indexer does:
Don't know the function name? semantic_search("validate email addresses") finds it anyway.
No grep. No file reading. No guessing. One query, exact location.
pip install symdex| Feature | One line | |
|---|---|---|
| Find | Symbol search | Locate any function, class, or method by name — returns exact byte offsets, no file reading |
| Find | Semantic search | Don't know the name? Search by what it does — semantic_search("validate email") finds it |
| Understand | Call graph | Who calls this function? What does it call? Pre-built at index time, instant at query time |
| Understand | HTTP route indexing | What routes does this API expose? search_routes answers without opening a single file |
| Stay current | Auto-watch | symdex watch — save a file, index updates automatically. Delete a file, it disappears from the index. No manual steps. |
| Scale | Cross-repo registry | One SymDex, many projects. Search across all indexed repos simultaneously |
| Clean up | Stale index GC | symdex gc — running parallel agents in git worktrees? One command removes orphaned databases |
| Access | Full CLI | Every capability available from the terminal, no agent required |
| Access | 16 MCP tools | Every capability available to any MCP-compatible AI agent |
| Language support | 14 languages | Python · JS · TS · Go · Rust · Java · PHP · C# · C · C++ · Elixir · Ruby · Vue + more via tree-sitter |
Every time an AI coding agent needs to find a function, it reads the entire file that might contain it:
Agent thought: "I need to find the validate_email function."
Agent action: Read auth/utils.py → 7,500 tokens consumed
Agent action: Read auth/validators.py → 6,200 tokens consumed
Agent action: Read core/helpers.py → 8,100 tokens consumed
Agent finds it on the third try. → 21,800 tokens wasted
This is the equivalent of reading an entire book from page one every time you want to find a single paragraph — when the book has an index sitting right there.
On a large codebase, a single development session can burn hundreds of thousands of tokens this way. That is real money, real slowness, and real context-window pressure.
SymDex is the index.
Most code navigation tools solve one problem. SymDex solves three.
Every other tool — LSP, grep, symbol search — requires you to know what you are looking for. SymDex does not.
# You don't know what the function is called. You know what it does.
symdex semantic "check that a user's email address is properly formatted" --repo myproject┏━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Name ┃ Kind ┃ Score ┃ File ┃
┡━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━┩
│ validate_email │ function │ 0.91 │ auth/utils.py │
│ is_valid_address │ function │ 0.74 │ core/validators.py │
└──────────────────┴──────────┴───────────────────┴────────────────────────┘
Vector embeddings of every symbol's signature and docstring. Fully local — no API calls, nothing leaves your machine.
SymDex extracts HTTP routes from your source during indexing. No more opening route files to understand what an API exposes.
symdex routes myproject -m POST
# → POST /users → create_user (api/views.py)# → POST /auth/login → login_user (auth/views.py)Supports Flask, FastAPI, Django, and Express. Agents can call search_routes directly via MCP.
symdex watch ./myproject # index now, then reindex on every saveSave a file — SymDex reindexes it. Delete a file — SymDex removes it from the index. The agent always sees the current state of your code without you doing anything.
┌─────────────────────────────────────────────────────────────────┐
│ STEP 1 — Index once (you run this, takes seconds to minutes) │
│ │
│ symdex index ./myproject │
│ │ │
│ ▼ │
│ tree-sitter parses every source file │
│ │ │
│ ▼ │
│ Every function, class, method extracted │
│ with name · kind · file · exact byte offsets · docstring │
│ │ │
│ ▼ │
│ Stored in SQLite database + vector embeddings (sqlite-vec) │
└─────────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────────┐
│ STEP 2 — Agent queries SymDex instead of reading files │
│ │
│ Without SymDex: │
│ Agent → read auth/utils.py (full) → 7,500 tokens │
│ │
│ With SymDex: │
│ Agent → search_symbols("validate_email") │
│ → { file: "auth/utils.py", start_byte: 1024, │
│ end_byte: 1340 } → ~200 tokens │
│ Agent → read bytes 1024–1340 only → done │
└─────────────────────────────────────────────────────────────────┘
SymDex does not read files for the agent. It tells the agent exactly where to look — file path and byte offset — so the agent reads only the bytes it needs. Nothing more.
Setup — index the project once:
symdex index ./myproject --name myproject
symdex serve # start the MCP serverAgent calls search_symbols to locate a function:
// Tool call
{ "tool": "search_symbols", "query": "validate_email", "repo": "myproject" }
// Response (~200 tokens)
{
"symbols": [
{
"name": "validate_email",
"kind": "function",
"file": "auth/utils.py",
"start_byte": 1024,
"end_byte": 1340,
"signature": "def validate_email(email: str) -> bool"
}
]
}Agent calls get_symbol to read only that function:
// Tool call — reads bytes 1024 to 1340 only
{ "tool": "get_symbol", "file": "auth/utils.py", "start_byte": 1024, "end_byte": 1340, "repo": "myproject" }
// Response — the exact function source, nothing else
{
"source": "def validate_email(email: str) -> bool:\n\"\"\"Validate email format.\"\"\"\n pattern = r'^[\\w.-]+@[\\w.-]+\\.\\w+$'\n return bool(re.match(pattern, email))"
}Agent calls get_callers to understand impact before changing it:
{ "tool": "get_callers", "name": "validate_email", "repo": "myproject" }
{
"callers": [
{ "name": "register_user", "file": "auth/views.py", "kind": "function" },
{ "name": "update_profile", "file": "users/views.py", "kind": "function" }
]
}Agent uses semantic_search when it doesn't know the exact name:
{ "tool": "semantic_search", "query": "check if user email address is valid", "repo": "myproject" }
// Finds by meaning, not by name
{
"symbols": [
{ "name": "validate_email", "score": 0.91, "file": "auth/utils.py" },
{ "name": "is_valid_address", "score": 0.74, "file": "core/validators.py" }
]
}Total tokens for this entire session: ~800 tokens. Without SymDex, finding and reading these three functions would cost ~25,000 tokens.
Here is how SymDex compares to the tools people reach for:
| Capability | LSP (pylsp, tsserver) | CodeGraphContext | Serena | SymDex |
|---|---|---|---|---|
| Find symbol by name | Yes | Yes | Yes | Yes |
| Search by meaning / intent | No | No | No | Yes — semantic_search |
| HTTP route indexing | No | No | No | Yes — search_routes |
| Auto-watch, live reindex | No | No | No | Yes — symdex watch |
| Call graph | Partial | Yes | Yes | Yes |
| Cross-repo search | No | No | No | Yes |
| Works without an editor | No | No | No | Yes — terminal-native |
| Zero-config storage | No | No (graph DB) | No | Yes — one SQLite file |
| pip install and done | No | No | No | Yes |
| Requires language server per language | Yes | No | Yes | No |
| Works offline (no API calls) | Yes | Yes | Yes | Yes |
LSP servers (pylsp, typescript-language-server, rust-analyzer) are excellent for editors. They require a running editor process, a language server installed and running per language, and they operate on live files in real time.
SymDex is terminal-native and editor-free. An agent running in a terminal (Claude Code, Codex CLI, OpenCode) gets the same symbol lookup capability with zero editor dependency. And LSP cannot do semantic search — if you don't know the function name, LSP cannot help.
CodeGraphContext builds a graph database over your code. The tradeoff: you need to choose and run a graph database backend (KùzuDB, Neo4j). SymDex uses SQLite — one file per repo, zero configuration. No backend, no server, no Docker.
CodeGraphContext has no semantic search and no HTTP route indexing. SymDex has both.
Serena wraps real language servers and gets true type-aware analysis — it can resolve which concrete implementation is called through an interface, track generics, follow pointer dispatch. That is genuinely powerful for large, strongly-typed codebases.
The tradeoff: Serena requires language servers installed per language and queries hit live files rather than a pre-built index. SymDex is faster per query (pre-indexed), works offline, and adds semantic search and route indexing — capabilities no language server provides.
If you are an AI agent reading this: Follow these rules to use SymDex correctly.
Rule 1 — Index before anything else.
Call index_folder with the project path at the start of every session. Without this, the index is empty and all searches return nothing.
{ "tool": "index_folder", "path": "/path/to/project", "name": "myproject" }Rule 2 — Never read a full file to find a function.
Always call search_symbols first. Use the returned start_byte and end_byte to read only the relevant bytes via get_symbol.
Rule 3 — Use the right tool for the right query.
| What you want | Tool to call |
|---|---|
| Find a function by name | search_symbols |
| Read a function's source | get_symbol (use byte offsets from search) |
| Find by meaning / description | semantic_search |
| See all symbols in a file | get_file_outline |
| Understand project structure | get_repo_outline |
| Find who calls a function | get_callers |
| Find what a function calls | get_callees |
| Search for a string in code | search_text |
| List all HTTP routes | search_routes |
Rule 4 — Re-index after code changes.
Call index_folder again (or invalidate_cache for a specific file) after modifying source files so the index reflects the latest state.
| Capability | Conventional (read files) | SymDex |
|---|---|---|
| Find a function by name | Read entire file(s) | Byte-offset lookup — read only those bytes |
| Token cost per lookup | ~7,500 tokens (one file) | ~200 tokens |
| Token cost across a session | Compounds per lookup | Fixed per lookup — does not compound |
| Search by meaning | Not possible | Semantic embedding search — finds by intent |
| "Who calls this function?" | Read every file manually | Pre-built call graph — instant answer |
| "What does this function call?" | Read function body manually | Pre-built call graph — instant answer |
| "What API routes does this repo expose?" | Read every route file | search_routes — instant, no file reading |
| Search across multiple projects | Not possible | Cross-repo registry — one SymDex, many projects |
| Keep index current after edits | Manual re-run | symdex watch — auto-reindex on save |
| Context window pressure | High — full files accumulate | Low — precise snippets only |
| Works with any AI agent | Agent-specific plugins | Any MCP-compatible agent — one config |
| Requires editor / language server | Often yes | No — standalone, terminal-native |
| Command-line access | Not available | Full CLI included |
| Re-index on changes | Full re-read every time | SHA-256 change detection — only re-indexes changed files |
Find code by what it does, not what it is called. SymDex embeds every symbol's signature and docstring into a vector and finds the closest matches by meaning — not by keyword. Powered by sentence-transformers running fully locally, no API calls required.
symdex semantic "parse and validate an authentication token" --repo myprojectFind any function, class, method, or variable by name across your entire indexed codebase. Returns file path and exact byte offsets. No file reading required.
Understand the impact of any change before you make it.
symdex callers process_payment --repo myproject # Who calls this? (impact analysis)
symdex callees process_payment --repo myproject # What does this call? (dependency trace)Call relationships are extracted during indexing and stored as a graph. No file reading at query time.
Run symdex watch once. Every time you save a file, SymDex automatically re-indexes only the changed file. Delete a file — SymDex removes it from the index. The agent always sees the current state of your code.
symdex watch ./myproject # Index now, then watch for changes
symdex watch ./myproject --interval 3 # Check every 3 seconds (default: 5)Works as a background process alongside your development workflow. The index stays current without any agent interruption.
SymDex automatically extracts HTTP API routes during indexing and makes them searchable. No more reading route files to understand an API surface.
Supported frameworks: Flask · FastAPI · Django · Express
symdex routes myproject # All routes in the repo
symdex routes myproject -m POST # Only POST routes
symdex routes myproject -p /users # Routes matching a path patternVia MCP tool (agents can call this directly):
{ "tool": "search_routes", "repo": "myproject", "method": "GET" }
// → [{ "method": "GET", "path": "/users", "handler": "list_users", "file": "api/views.py" }, ...]Index multiple projects and search across all of them from one place.
symdex index ./frontend --name frontend
symdex index ./backend --name backend
symdex search "validate_token"# searches both repos simultaneouslyEach repo gets its own SQLite database. The registry tracks all of them.
SymDex stores a SHA-256 hash of every indexed file. Re-indexing only processes files that have actually changed. On large codebases this makes incremental updates take seconds, not minutes.
Every MCP tool is also available as a CLI command. Use SymDex without an AI agent — in scripts, in CI, or just to explore your codebase.
Run SymDex as a local stdio server (default, for desktop agents) or as an HTTP server for remote access.
symdex serve # stdio — for Claude, Cursor, Copilot, Gemini CLI, Codex CLI, etc.
symdex serve --port 8080 # HTTP — for remote agents or servicesSymDex parses source files using tree-sitter — a fast, robust, incremental parser used by major editors including Neovim, Helix, and GitHub.
| Language | File Extensions |
|---|---|
| Python | .py |
| JavaScript | .js.mjs |
| TypeScript | .ts.tsx |
| Go | .go |
| Rust | .rs |
| Java | .java |
| PHP | .php |
| C# | .cs |
| C | .c.h |
| C++ | .cpp.cc.h |
| Elixir | .ex.exs |
| Ruby | .rb |
| Vue | .vue |
13 languages. More can be added by installing additional tree-sitter grammar packages.
SymDex speaks the Model Context Protocol (MCP) — the open standard for connecting AI agents to external tools. If a platform supports MCP, SymDex works with it — no custom integration required.
| Platform | By | How to Connect |
|---|---|---|
| Claude Desktop | Anthropic | Add to claude_desktop_config.json |
| Claude Code | Anthropic | claude mcp add symdex -- symdex serve |
| Codex CLI | OpenAI | Add to MCP settings |
| Codex App | OpenAI | Add to MCP settings |
| Gemini CLI | Add to MCP settings | |
| Cursor | Anysphere | Add to .cursor/mcp.json |
| Windsurf | Codeium | Add to MCP settings |
| GitHub Copilot (agent mode) | Microsoft | Add to .vscode/mcp.json |
| Continue.dev | Continue | Add to config.json |
| Cline | Cline | Add to MCP settings |
| Kilo Code | Kilo Code | Add to VS Code settings.json under kilocode.mcpServers |
| Zed | Zed Industries | Add to MCP settings |
| OpenCode | OpenCode | Add to opencode.json |
| Any custom MCP client | — | stdio or HTTP transport |
{
"mcpServers": {
"symdex": {
"command": "symdex",
"args": ["serve"]
}
}
}For HTTP mode (remote agents):
{
"mcpServers": {
"symdex": {
"url": "http://localhost:8080/mcp"
}
}
}Available on PyPI:
pip install symdexRequires Python 3.11 or higher.
symdex index ./myproject --name myprojectSymDex walks the directory, parses every supported source file, and writes the index to a local SQLite database. Run this once. Re-run it when your code changes (only modified files are re-processed).
symdex search "validate_email" --repo myproject┏━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┓
┃ Repo ┃ Kind ┃ Name ┃ File ┃ Start ┃
┡━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━┩
│ myproject │ function │ validate_email │ auth/utils.py │ 1024 │
└────────────────┴──────────┴────────────────┴─────────────────────────────────────────┴───────┘
symdex servePoint your agent at it using the config above. The agent can now use all 16 MCP tools.
These are the tools your AI agent can call once SymDex is running as an MCP server.
| Tool | Description |
|---|---|
index_folder | Index a local folder — run once per project |
index_repo | Index a named, registered repo |
search_symbols | Find function or class by name — returns byte offsets |
get_symbol | Retrieve one symbol's full source by byte offset |
get_symbols | Bulk symbol retrieval by a list of offsets |
get_file_outline | All symbols in a file — no file content transferred |
get_repo_outline | Directory structure and symbol statistics for a repo |
get_file_tree | Directory tree — structure only, no content |
search_text | Text or regex search — returns matching lines only |
list_repos | List all indexed repos in the registry |
invalidate_cache | Force re-index on next request |
semantic_search | Find symbols by meaning using embedding similarity |
get_callers | Find all functions that call a named function |
get_callees | Find all functions called by a named function |
search_routes | Find HTTP routes indexed from a repo (Flask/FastAPI/Django/Express) — filter by method or path |
gc_stale_indexes | Remove databases for repos whose directories no longer exist on disk |
# Indexing
symdex index ./myproject # Index a folder
symdex index ./myproject --name myproj # Index with a custom name
symdex invalidate --repo myproj # Force re-index a repo
symdex invalidate --repo myproj --file auth.py # Force re-index one file# Symbol search
symdex search "validate email" --repo myproj # Search by name
symdex search "validate email"# Search across all repos
symdex find MyClass --repo myproj # Exact name lookup# Semantic search
symdex semantic "authentication token parsing" --repo myproj
# File and repo inspection
symdex outline myproj/auth/utils.py --repo myproj # All symbols in a file
symdex repos # List all indexed repos
symdex text "TODO" --repo myproj # Text search# Call graph
symdex callers process_payment --repo myproj # Who calls this function
symdex callees process_payment --repo myproj # What this function calls# Watch (auto-reindex on file changes)
symdex watch ./myproject # Auto-reindex on file changes
symdex watch ./myproject --interval 10 # Custom poll interval (seconds)# Routes
symdex routes myproject # List all indexed HTTP routes
symdex routes myproject -m GET # Filter by HTTP method# Maintenance
symdex gc # Remove stale indexes for deleted/moved repos# Server
symdex serve # Start MCP server (stdio)
symdex serve --port 8080 # Start MCP server (HTTP)Click to expand — internals for the technically curious
Each indexed repo gets its own SQLite database file stored in ~/.symdex/. A shared registry database tracks all repos.
-- Every extracted symbol
symbols (
id INTEGERPRIMARY KEY,
repo TEXTNOT NULL,
file TEXTNOT NULL,
name TEXTNOT NULL,
kind TEXTNOT NULL, -- function | class | method | constant | variable
start_byte INTEGERNOT NULL,
end_byte INTEGERNOT NULL,
signature TEXT,
docstring TEXT,
embedding BLOB -- float32 vector stored via sqlite-vec
)
-- Call graph edges
edges (
caller_id INTEGERREFERENCES symbols(id),
callee_name TEXTNOT NULL,
callee_file TEXT
)
-- Change detection
files (
repo TEXTNOT NULL,
pathTEXTNOT NULL,
hash TEXTNOT NULL, -- SHA-256 of file contents
indexed_at DATETIME NOT NULL,
PRIMARY KEY (repo, path)
)
-- Cross-repo registry
repos (
name TEXTPRIMARY KEY,
root_path TEXTNOT NULL,
db_path TEXTNOT NULL,
last_indexed DATETIME
)
-- HTTP routes (Flask, FastAPI, Django, Express)
routes (
repo TEXTNOT NULL,
file TEXTNOT NULL,
method TEXTNOT NULL, -- GET | POST | PUT | DELETE | PATCH | ANYpathTEXTNOT NULL, -- /users/{id}
handler TEXT, -- function name
start_byte INTEGERNOT NULL,
end_byte INTEGERNOT NULL
)Source files are parsed using tree-sitter. tree-sitter produces a concrete syntax tree for each file. SymDex walks the tree and extracts nodes matching known symbol types per language (e.g. function_definition for Python, function_declaration for Go, method_definition for JavaScript).
When a symbol has a docstring or signature, SymDex generates a vector embedding using sentence-transformers (model: all-MiniLM-L6-v2 by default). Embeddings are stored as raw float32 blobs and queried using sqlite-vec — a SQLite extension for vector similarity search. Everything runs locally. No embedding API calls.
Built on FastMCP. Supports both stdio transport (for desktop agents) and streamable HTTP transport (for remote access).
symdex/
├── cli.py — Typer CLI (all user-facing commands)
├── core/
│ ├── parser.py — tree-sitter symbol extraction (13 languages + Vue)
│ ├── storage.py — SQLite read/write, vector storage, route storage
│ ├── indexer.py — orchestrates parse → store pipeline
│ ├── watcher.py — file-system watcher (watchdog), auto-reindex
│ ├── route_extractor.py — regex-based HTTP route detection
│ └── schema.sql — database schema (symbols, edges, files, repos, routes)
├── mcp/
│ ├── server.py — FastMCP server definition
│ └── tools.py — 16 MCP tool implementations
├── search/
│ ├── symbol_search.py — name-based FTS search
│ ├── text_search.py — regex/text search
│ └── semantic.py — embedding similarity search
└── graph/
├── call_graph.py — call edge extraction and query
└── registry.py — cross-repo registry and multi-DB search
Do I need to re-index every time I change my code?
No. Run symdex watch ./myproject once and the index stays current automatically — every file save triggers a reindex of that file, every delete removes it from the index. If you prefer manual control, run symdex index again at any point — SHA-256 hashing means only files that actually changed are reprocessed.
Does semantic search send my code to an API?
No. Embeddings are generated locally using sentence-transformers. Nothing leaves your machine.
Can I use SymDex without an AI agent? Yes. The full CLI gives you direct access to every search capability — symbol search, semantic search, call graph, file outlines — without any agent involved.
Does it work with monorepos?
Yes. Index each sub-project separately with a unique --name, then search across all of them using symdex search without a --repo flag.
Does it handle pointers, generics, and type-resolution?
SymDex extracts symbols, signatures, and docstrings from parsed syntax trees. It does not perform full type inference or resolve generics at the type system level — that requires a language server per language (pylsp, rust-analyzer, etc.). SymDex's semantic search compensates: if you don't know the exact name of a generic method, semantic_search finds it by meaning.
What happens if a language is not supported? SymDex skips files with unrecognised extensions. Supported and unsupported files can coexist in the same project — only the supported ones are indexed.
Is the index portable?
Yes. The SQLite .db files can be copied to another machine. As long as SymDex is installed there, the index works. The only caveat is that absolute file paths in the index will point to the original machine.
MIT — see LICENSE
Issues and pull requests are welcome at github.com/husnainpk/SymDex.