SymDex

Repo-local code intelligence for AI coding agents.

Index a project once, then give agents exact symbols, routes, callers, callees, file outlines, and semantic matches instead of whole files.

7,500 tokens -> 200 tokens

Per lookup. Every lookup. Approximate, but directionally real.

# Install the lean core
pip install symdex

# Add local semantic search when you want the sentence-transformers backend
pip install "symdex[local]"

# Or install the local semantic backend as an isolated CLI tool
uv tool install "symdex[local]"

# Or run without installing
uvx symdex --help

# Upgrade an existing install
py -m pip install -U symdex
uv tool upgrade symdex
uvx symdex@latest --help

# Install the SymDex agent skill globally for supported agents
npx skills add https://github.com/husnainpk/SymDex --skill symdex-code-search --yes --global

---

What SymDex Does

AI coding agents are useful until they have to rediscover your repo from scratch. They open whole files, grep broad patterns, miss the route handler, read the same utility twice, and spend thousands of tokens just getting oriented.

SymDex gives agents a repo-local retrieval layer before they start reading code. You keep the index on your machine, choose the embedding backend only when semantic search is useful, and expose the same project map through both CLI commands and MCP tools.

It indexes your project into a small local SQLite knowledge base with:

• exact symbols and byte offsets
• file outlines and repo summaries
• literal text search
• optional semantic search
• callers, callees, and circular dependency checks
• extracted HTTP routes
• a registry for one repo, many repos, or many worktrees

Then agents can ask for the narrow slice they need: the function, route, caller chain, file outline, or intent match. That means less blind browsing, less context waste, and responses that can explain how much token budget SymDex saved.

SymDex is local-first. Base symdex keeps symbol, text, route, graph, and MCP features lean. Install symdex[local] only when you want local semantic embeddings, or point the hosted backend at Voyage, OpenAI-compatible services, Gemini, or a compatible proxy when you want remote embeddings.

Current release capabilities:

• current release 0.1.25; latest public tag v0.1.25
• 20 MCP tools across indexing, search, outlines, routes, stats, graphs, cache invalidation, and stale-index cleanup
• 21 language surfaces, including HTML, CSS, Shell, Svelte, Markdown headings, and supported fenced code blocks
• Android, Flutter, and iOS coverage through Kotlin, Dart, and Swift parser targets
• route extraction across Python, JavaScript/TypeScript, Spring/Kotlin, Laravel, Gin-style Go, ASP.NET, Rails/Sinatra, Phoenix, and Actix
• one-line CLI token-savings footers after successful search commands
• MCP roi, roi_summary, and roi_agent_hint fields so agents can mention savings in their final response
• semantic backends for local sentence-transformers, Voyage, OpenAI-compatible /embeddings, and Gemini Embedding
• SYMDEX_EMBED_RPM request pacing for hosted embedding providers
• symdex index --lazy for fast foreground structural indexing while embeddings build in a background watcher
• low-memory symdex watch by default: structural refresh without loading embedding models unless --embed is passed
• duplicate watcher protection, idle auto-exit, and state-aware watcher metadata
• workspace-local ./.symdex state for Docker, portable workspaces, and teams that do not want indexes hidden in home directories
• upgrade notices with exact pip, uv tool, and uvx commands when a newer release exists

---

SymDex Skill For Agents

Install the SymDex code-search skill when you want supported agents to use SymDex before broad file browsing:

npx skills add https://github.com/husnainpk/SymDex --skill symdex-code-search --yes --global

If you want the interactive installer instead, omit --yes --global.

The skill tells agents to:

• check repo and index readiness first
• search with SymDex before broad Read/Grep/Glob discovery
• prefer symbol-level and outline-level retrieval over full-file reads
• use callers, callees, routes, repo stats, and semantic search when those are the better fit
• fall back clearly when SymDex is unavailable or when semantic embeddings have not been built
• keep the workflow centered on lower-token code retrieval instead of broad file reads

The skill lives in this repo at skills/symdex-code-search/SKILL.md and follows the standard skills/<name>/SKILL.md layout.

Installing through the skills CLI uses the same public path that skills.sh indexes.

---

60-second quickstart

# Install the lean core
pip install symdex
# or
uv tool install symdex

# Index a project
symdex index ./myproject --repo myproject

# Search by symbol name
symdex search "validate_email" --repo myproject

# Search by literal text
symdex text "JWT" --repo myproject

# Add local semantic search only when you want embeddings
pip install "symdex[local]"

# Search by intent
symdex semantic "check email format" --repo myproject

# Show HTTP routes
symdex routes myproject -m POST

# Start the MCP server
symdex serve

Notes:

• If you omit --repo on symdex index or symdex watch, SymDex auto-generates a stable repo id from the current git branch and worktree path hash.
• After indexing, SymDex prints a code summary.
• After successful search commands, SymDex prints a one-line ROI footer with approximate token savings.
• MCP search tools also return roi_agent_hint, so agents can fold the savings into their user-facing response instead of burying it in logs.
• When a newer PyPI release exists, normal CLI commands print exact upgrade commands for pip, uv tool, and uvx.
• Set SYMDEX_STATE_DIR=.symdex on first index to keep repo databases, registry.db, and registry.json inside the current workspace. After that, commands run from the workspace auto-discover the local state.
• --state-dir can be passed either globally or after the subcommand, for example symdex --state-dir .symdex repos or symdex repos --state-dir .symdex.
• Canonical CLI commands are index and repos. Shell compatibility aliases also accept MCP-shaped names like index-folder, index-repo, and list-repos.
• Semantic search requires stored embeddings. If a repo was indexed before the backend you want was enabled, re-index it with symdex index, symdex index --lazy, or symdex watch --embed.

Add to your agent config:

{
  "mcpServers": {
    "symdex": {
      "command": "uvx",
      "args": ["symdex", "serve"]
    }
  }
}

HTTP mode:

{
  "mcpServers": {
    "symdex": {
      "url": "http://localhost:8080/mcp"
    }
  }
}

---

Workspace-local state and Docker

SymDex defaults to ~/.symdex and also supports a workspace-local state directory for portable and containerized workflows.

Use it like this on first setup:

SYMDEX_STATE_DIR=.symdex symdex index ./myproject --repo myproject

That creates:

• ./.symdex/<repo>.db
• ./.symdex/registry.db
• ./.symdex/registry.json

registry.json is the human-readable manifest. In workspace-local mode it stores relative root_path and db_path values, so you can inspect what is indexed without opening SQLite.

After the local state exists, SymDex auto-discovers it from the current workspace or any nested subdirectory.

---

What you get

Feature	Details
Symbol search	Find functions, classes, and methods with exact byte offsets
Semantic search	Find code by intent instead of exact name
Text search	Search indexed files by literal text across the repo
Byte-precise retrieval	Read only the symbol span you need
File outline	List symbols in a file without transferring the whole file
Repo outline	Get a directory tree plus repo summary through MCP
Call graph	Trace callers, callees, and circular dependencies
HTTP routes	Extract Flask, FastAPI, Django, Express, Spring/Kotlin, Laravel, Gin-style Go, ASP.NET, Rails/Sinatra, Phoenix, and Actix routes
Markdown-aware docs	Index Markdown headings and supported fenced code blocks for SDK docs, specs, repo docs, and design notes
Auto-watch	Re-index on change, avoid duplicate watchers, stay low-memory by default, and auto-exit after idle time
Cross-repo registry	Manage multiple indexed repos from one local registry
Workspace-local state	Keep repo databases, registry.json, and watcher metadata inside ./.symdex for Docker and portable workspaces
Search ROI footer	One-line approximate token savings after successful search commands
Agent ROI hint	MCP tools return roi_agent_hint so agents can mention savings naturally in their replies
Code summary	Files, Lines of Code, symbols, routes, skipped files, and languages after indexing
Optional embedding backends	Use local embeddings, Voyage, OpenAI-compatible /embeddings, Gemini, or a compatible proxy only when needed

---

Where SymDex Fits

AI codebase tools now split into a few strong categories: enterprise code search, IDE-native assistants, LSP-backed agent toolkits, repo-map systems, and whole-repo prompt packers. SymDex sits in a narrower but useful slot: a local-first retrieval layer that combines exact symbols, text search, optional semantic search, HTTP routes, callers, callees, file outlines, repo outlines, Markdown/fenced-code indexing, repo stats, token-savings hints, CLI access, and MCP access without requiring an editor session or hosted index.

That makes SymDex strongest when an agent needs a portable codebase map it can query before reading files.

Need	Common fit	Where SymDex fits
Enterprise code search across many hosted repos	Sourcegraph-style code search	SymDex is smaller and local-first; it focuses on agent retrieval from checked-out worktrees rather than enterprise code search operations
IDE-native codebase chat and autocomplete	Cursor, Continue, Cody, and similar editor assistants	SymDex is editor-agnostic; agents can use the same repo index from a terminal, stdio MCP, or HTTP MCP client
IDE-grade semantic retrieval and refactoring	LSP-backed tools and Serena-style agent toolkits	SymDex does not try to be a refactoring engine; it focuses on retrieval primitives agents need before deciding what to edit
Whole-repo context packaging	Repomix and other prompt-packing tools	SymDex retrieves targeted slices instead of packing the whole repository into one prompt artifact
Agent repo maps for coding sessions	Aider-style repository maps	SymDex exposes queryable indexes, routes, call graphs, Markdown headings, semantic search, and ROI hints rather than only a condensed map
Lightweight local retrieval across code and docs	SymDex	SymDex's edge is the combined surface: CLI + MCP, SQLite indexes, 21 language surfaces, route extraction, call graphs, Markdown/MDX support, optional embeddings, low-memory watch, and workspace-local state

Use a language server when you need deep type-system reasoning or editor-native refactors. Use an enterprise code search platform when you need organization-wide search, permissions, and code monitoring. Use a prompt packer when you want to hand an entire repo snapshot to a model. Use SymDex when you want an AI coding agent to ask precise repo-local questions before it spends tokens reading files.

---

CLI reference

# Indexing and maintenance
symdex index ./myproject --repo myproject       # Index with an explicit repo id
symdex index ./myproject                        # Auto-name from git branch + path hash
symdex index ./myproject --repo myproject --no-embed  # Skip semantic embedding work
symdex index ./myproject --repo myproject --lazy      # Index code now, embed in a background watcher
symdex watch ./myproject --repo myproject       # Keep an index fresh without loading embedding models
symdex watch ./myproject --repo myproject --embed  # Also refresh semantic embeddings on change
symdex invalidate --repo myproject              # Force re-index of a repo
symdex invalidate --repo myproject --file app.py
symdex gc                                       # Remove stale index databases
symdex repos                                    # List indexed repos

# Search
symdex search "validate_email" --repo myproject
symdex search "validate_email"                 # Search across all indexed repos
symdex find validate_email --repo myproject     # Exact lookup
symdex text "JWT" --repo myproject
symdex semantic "check auth token" --repo myproject

# Navigation
symdex outline auth/utils.py --repo myproject
symdex callers validate_email --repo myproject
symdex callees validate_email --repo myproject
symdex routes myproject
symdex routes myproject --method POST
symdex routes myproject --path /api

# Server
symdex serve
symdex serve --port 8080

MCP currently exposes additional repo-tree and repo-stats views that are not surfaced as dedicated CLI commands: get_file_tree, get_repo_outline, get_index_status, and get_repo_stats.

---

MCP tools

SymDex currently exposes 20 MCP tools:

Tool	Purpose
index_folder	Index a local folder and return indexing statistics
index_repo	Index a repo and register it in the central registry
search_symbols	Find functions, classes, and methods by name
semantic_search	Find symbols by meaning using embedding similarity
search_text	Search indexed files by text and return matching lines
get_symbol	Read one symbol by byte offsets
get_symbols	Bulk exact-name symbol lookup
get_file_outline	List symbols in a single file
get_file_tree	Return a directory tree without file contents
get_repo_outline	Return a repo tree plus summary stats
get_callers	Return symbols that call a named function
get_callees	Return symbols called by a named function
search_routes	Query extracted HTTP routes
get_index_status	Return symbol count, file count, Lines of Code, staleness, and watcher state
get_repo_stats	Return repo metrics such as language mix, fan-in, fan-out, and circular dependency count
get_graph_diagram	Generate a Mermaid call graph
find_circular_deps	Detect circular dependencies
list_repos	List all indexed repos
invalidate_cache	Force re-index on next use
gc_stale_indexes	Remove index databases for repos that no longer exist on disk

---

Supported languages

Language	Extensions
Python	.py
JavaScript	.js, .jsx, .mjs, .cjs, .cjsx, .mjsx
TypeScript	.ts, .tsx, .mts, .cts, .mtsx, .ctsx
Go	.go
Rust	.rs
Java	.java
Kotlin	.kt, .kts
Dart	.dart
Swift	.swift
PHP	.php
C#	.cs
C	.c
C++	.h, .hh, .hpp, .hxx, .cpp, .cc, .cxx
HTML	.html, .htm
CSS	.css, .scss, .sass, .less, .stylus, .styl
Shell	.sh, .bash, .zsh
Elixir	.ex, .exs
Ruby	.rb
Vue	.vue script blocks parsed as JavaScript or TypeScript
Svelte	.svelte script blocks parsed as JavaScript or TypeScript
Markdown	.md, .markdown, .mdx headings plus supported fenced code blocks

Powered by tree-sitter for code, language-pack grammar fallbacks, and a native Markdown scanner for headings and fenced code examples.

---

Supported platforms

SymDex works with any MCP client that supports stdio or streamable HTTP.

Platform	Typical setup
Codex CLI	Add to MCP settings
Gemini CLI	Add to MCP settings
Cursor	.cursor/mcp.json
Windsurf	Add to MCP settings
GitHub Copilot	.vscode/mcp.json
Roo	Add to MCP settings
Continue.dev	config.json
Cline	Add to MCP settings
Kilo Code	VS Code MCP settings
Zed	Add to MCP settings
OpenCode	opencode.json
Any MCP client	uvx symdex serve or symdex serve --port 8080

---

Semantic embedding backends

Base symdex installs the lean core only. Choose the embedding extra that matches how you want semantic search to work:

• symdex[local] for local sentence-transformers
• symdex[voyage] for Voyage text embeddings
• symdex[voyage-multimodal] for Voyage text plus images and PDFs

Remote OpenAI-compatible and Gemini backends use the Python standard library HTTP client, so they do not add a required dependency.

Local mode

pip install "symdex[local]"

SYMDEX_EMBED_BACKEND=local \
SYMDEX_EMBED_MODEL=all-MiniLM-L6-v2 \
symdex index . --repo myrepo

Voyage text mode

pip install "symdex[voyage]"

SYMDEX_EMBED_BACKEND=voyage \
VOYAGE_API_KEY=... \
SYMDEX_VOYAGE_MODEL=voyage-code-3 \
symdex index . --repo myrepo

SYMDEX_EMBED_BACKEND=voyage \
VOYAGE_API_KEY=... \
symdex semantic "parse source code" --repo myrepo

OpenAI-compatible mode

Use this for OpenAI, local OpenAI-compatible servers, and proxy services that expose POST /embeddings.

SYMDEX_EMBED_BACKEND=openai \
SYMDEX_EMBED_BASE_URL=https://api.openai.com/v1 \
SYMDEX_EMBED_MODEL=text-embedding-3-small \
SYMDEX_EMBED_API_KEY=... \
SYMDEX_EMBED_RPM=60 \
symdex index . --repo myrepo --lazy

For a local or proxy provider, set SYMDEX_EMBED_BACKEND=custom, point SYMDEX_EMBED_BASE_URL at its /v1 base URL, and set the model name it expects. If the provider does not need an API key, omit SYMDEX_EMBED_API_KEY.

Gemini mode

Gemini uses RETRIEVAL_DOCUMENT when indexing and RETRIEVAL_QUERY when searching.

SYMDEX_EMBED_BACKEND=gemini \
GEMINI_API_KEY=... \
SYMDEX_EMBED_MODEL=text-embedding-004 \
SYMDEX_EMBED_RPM=60 \
symdex index . --repo myrepo --lazy

Voyage multimodal mode

pip install "symdex[voyage-multimodal]"

SYMDEX_EMBED_BACKEND=voyage
SYMDEX_VOYAGE_MULTIMODAL=1
VOYAGE_API_KEY=...
SYMDEX_VOYAGE_MULTIMODAL_MODEL=voyage-multimodal-3.5
symdex index . --repo myrepo

Multimodal mode lets SymDex index supported images, screenshots, and PDFs as searchable asset entries.

Notes:

• Base symdex keeps symbol, text, route, and call-graph features without pulling in the local embedding stack.
• If SYMDEX_EMBED_BACKEND is unset, SymDex uses the local backend when symdex[local] is installed.
• Local semantic search requires symdex[local] and downloads the model on first use.
• Voyage text mode requires symdex[voyage].
• Voyage multimodal mode requires symdex[voyage-multimodal].
• OpenAI-compatible mode reads SYMDEX_EMBED_BASE_URL, SYMDEX_EMBED_MODEL, and optional SYMDEX_EMBED_API_KEY or OPENAI_API_KEY.
• Gemini mode reads SYMDEX_EMBED_MODEL, SYMDEX_EMBED_BASE_URL, and SYMDEX_EMBED_API_KEY, GEMINI_API_KEY, or GOOGLE_API_KEY.
• SYMDEX_EMBED_RPM applies request pacing to remote backends. Use it when free tiers or proxies enforce requests-per-minute limits.
• symdex index --lazy performs the fast structural index first, then starts a background symdex watch --embed process so embeddings can fill in without blocking the foreground command.
• symdex index --no-embed skips semantic embedding work entirely.
• If the selected backend extra is missing, SymDex prints an actionable install hint.
• Multimodal indexing is only active when SYMDEX_VOYAGE_MULTIMODAL=1.

---

FAQ

What is SymDex?

SymDex is a repo-local code intelligence tool for AI coding agents. It indexes a project into a local SQLite knowledge base so agents can retrieve exact symbols, file outlines, HTTP routes, callers, callees, text matches, semantic matches, and repo summaries without reading whole files first.

What problem does SymDex solve for AI coding agents?

SymDex reduces blind repo browsing. Instead of opening broad files and spending thousands of tokens on orientation, an agent can ask SymDex for the specific function, route handler, call graph edge, Markdown heading, or file outline needed for the current task.

Is SymDex an MCP server?

Yes. SymDex includes an MCP server with 20 tools for indexing, symbol search, semantic search, text search, file outlines, repo outlines, route search, call graph traversal, repo stats, cache invalidation, and stale-index cleanup. It also provides CLI commands for the same repo-local retrieval workflow.

Which AI coding agents can use SymDex?

Any MCP client that supports stdio or streamable HTTP can use SymDex. Typical setups include Codex CLI, Gemini CLI, Cursor, Windsurf, GitHub Copilot, Roo, Continue.dev, Cline, Kilo Code, Zed, OpenCode, and other MCP-compatible clients.

How is SymDex different from grep, ripgrep, embeddings, or an LSP?

SymDex combines exact structural search with optional semantic search. Grep and ripgrep find text; language servers provide editor-native type navigation; embedding-only systems find approximate meaning. SymDex gives agents a pre-indexed local map with symbols, byte offsets, routes, call graphs, text search, semantic search, and MCP tools in one workflow.

Does SymDex replace a language server?

No. SymDex is not a full type checker or automated refactoring engine. Use a language server when editor-native type reasoning or refactors matter. Use SymDex when an agent needs fast repo-local retrieval, route discovery, code outlines, semantic search, and lower-token navigation across one or more repos.

What languages and files does SymDex index?

SymDex currently covers 21 language surfaces: Python, JavaScript, TypeScript, Go, Rust, Java, Kotlin, Dart, Swift, PHP, C#, C, C++, HTML, CSS-family stylesheets, Shell, Elixir, Ruby, Vue script blocks, Svelte script blocks, and Markdown. Markdown support includes .md, .markdown, and .mdx headings plus supported fenced code blocks.

Can SymDex extract HTTP routes?

Yes. SymDex extracts HTTP routes across Python, JavaScript/TypeScript, Spring/Kotlin, Laravel, Gin-style Go, ASP.NET, Rails/Sinatra, Phoenix, and Actix. Agents can query routes directly instead of manually scanning routers, decorators, controllers, or framework entry points.

Does SymDex help documentation, SEO, or Generative Engine Optimization teams?

Yes, when technical content depends on a codebase. Documentation, SEO, and Generative Engine Optimization (GEO) teams can use SymDex-powered agents to find SDK examples, API routes, Markdown headings, feature implementations, and source-backed answers before updating developer docs, changelogs, comparison pages, or AI-search-friendly technical content. SymDex is a code retrieval layer, not a keyword research or rank tracking tool.

Does semantic search require the internet?

No, not by default. Install symdex[local] for local sentence-transformers semantic search; the model downloads once and then runs offline. Voyage, OpenAI-compatible, and Gemini embedding modes require network access unless the configured base URL points at a local server or private proxy.

Can SymDex run without semantic embeddings?

Yes. pip install symdex keeps the core symbol, text, route, file outline, repo outline, call graph, and MCP features lean. Use symdex index --no-embed to skip embedding work entirely, or add embeddings later with symdex[local], Voyage, OpenAI-compatible, Gemini, symdex index --lazy, or symdex watch --embed.

Why does symdex semantic say a repo has no semantic embeddings?

That repo was indexed without an embedding backend, indexed with --no-embed, or kept fresh by the default low-memory watch mode. Enable the backend you want, then run symdex index, symdex index --lazy, or symdex watch --embed so semantic vectors are written into the index.

Will symdex watch keep a large embedding model in memory?

No, not by default. symdex watch refreshes structural indexes without loading local embedding models unless --embed is passed. Watch mode also refuses duplicate watchers for the same repo/root, can auto-exit after idle time, and stores watcher metadata in the active state directory.

Where are SymDex indexes stored?

By default, each repo gets its own SQLite database under ~/.symdex, plus a central registry database. Set SYMDEX_STATE_DIR=.symdex or use symdex --state-dir .symdex ... to keep repo databases, registry.db, registry.json, and watcher metadata inside the current workspace.

Can SymDex work across multiple repos and worktrees?

Yes. SymDex maintains a registry of indexed repos, supports explicit --repo names, and can auto-generate stable repo ids from the current branch and worktree path when --repo is omitted. Use symdex repos to list indexes and symdex gc to clean stale entries after deleting worktrees.

How does SymDex report token savings?

Successful CLI search commands print a one-line ROI footer with approximate token savings. MCP search tools return structured roi, roi_summary, and roi_agent_hint fields so agents can mention retrieval savings in user-facing responses.

What does indexing print?

After indexing, SymDex prints a code summary with file count, Lines of Code, symbol counts, route counts, skipped files, errors, and language breakdown. This helps teams confirm what was indexed before relying on search results.

Can generated files be excluded?

Yes. Add a .symdexignore file at the repo root with gitignore-style patterns. SymDex also skips common generated, dependency, and build paths by default.

Why do some logs say index_folder or list_repos while CLI docs say index and repos?

index_folder and list_repos are MCP tool names. The canonical shell commands are symdex index and symdex repos, and compatibility aliases such as symdex index-folder and symdex list-repos are also accepted.

Do existing users get update notices?

Yes. Interactive CLI commands can print a brief upgrade notice with exact commands for pip, uv tool, and uvx. JSON output stays quiet so structured consumers are not broken.

Can SymDex be used without an AI agent?

Yes. All core search and navigation features are available through the CLI, including symbol search, exact lookup, text search, semantic search, route search, callers, callees, outlines, repo listing, and stale-index cleanup.

---

Contributing

Issues and PRs are welcome at github.com/husnainpk/SymDex.

If SymDex saves you tokens, a star helps other people find it.