Reflex

Sub-100ms local code search — CLI, scripts, and AI agents

Reflex is a local-first, full-text code search engine. Use it from the command line, pipe it into scripts, or connect it to AI coding assistants (Claude Code, Cursor, and any MCP-compatible tool) for instant symbol lookup, dependency analysis, and codebase exploration — fully offline, fully deterministic, no cloud required.

Quick start

1. Install

# Via NPM
npm install -g reflex-search

# Or via Cargo
cargo install reflex-search

2. Index and search

# From your project root
rfx index

# Full-text search
rfx query "extract_symbols"

# Symbol definitions only
rfx query "CacheManager" --symbols

# JSON output for scripting
rfx query "TODO" --json --limit 20

3. (Optional) Connect to an AI agent via MCP

Add this to your Claude Code MCP configuration (~/.claude/claude_desktop_config.json):

{
  "mcpServers": {
    "reflex": {
      "command": "rfx",
      "args": ["mcp"]
    }
  }
}

Your AI assistant can now call search_code, find_references, get_dependencies, and more.

See Claude Code + Reflex MCP Quickstart for MCP setup, key tools, and troubleshooting.

Why Reflex vs. built-in search tools

Measured efficiency (A/B vs. built-in AI search)

We A/B-tested an AI coding agent on real code-search tasks using Reflex (via MCP) against the same agent using its built-in search (ripgrep-backed Grep/Glob) — identical tasks, model, and repository, paired per task. The harness lives in benches/efficacy/ and is fully reproducible.

Setup: model claude-sonnet-4-6; 12 code-search tasks (find-all-usages, symbol locate, dependency/reverse-dependency, hotspot, comprehension, plus negative controls); N = 3 replicates per arm; run against the Reflex repository.

Results — Reflex ÷ built-in, so < 1.0 means Reflex uses less:

Implications

• No-regret replacement for built-in search. Reflex matches built-in tools on answer correctness (100% task success in both arms) at parity-or-better token usage and meaningfully lower dollar cost.
• Fewer round-trips on navigation. find_references returns a symbol's definition and every call site in one call, so the agent iterates less than chaining grep + file reads.
• The gap should widen with repo size. The baseline here is ripgrep — already fast on a mid-size repo. Reflex's trigram index is built to win most where linear scans are slowest: very large codebases and whole-repo "find every occurrence" tasks.

Honest caveats. This is a focused benchmark: one model, one repository, N = 3 — enough to demonstrate parity-to-better and no regression, not a large statistical claim (the token primary is formally "no significant difference," with point estimates favoring Reflex). Per-result precision/recall is not yet formally scored. Reproduce it yourself:

python3 benches/efficacy/runner.py --arms A B --repos reflex --n 3
python3 benches/efficacy/extract_metrics.py && python3 benches/efficacy/analyze.py

MCP tools

When connected via MCP, your AI assistant gets these tools:

Index not found error? If an MCP tool returns "Index not found. Run 'rfx index' to build the cache first", call index_project first, then retry the failed tool.

CLI usage

Reflex also works as a standalone CLI for humans and shell scripts.

# Full-text search (finds every occurrence)
rfx query "extract_symbols"

# Symbol definitions only (faster, uses tree-sitter)
rfx query "extract_symbols" --symbols

# Filter by language and symbol kind
rfx query "parse" --lang rust --kind function --symbols

# Regex search
rfx query "fn.*test" --regex

# JSON output for programmatic use
rfx query "unwrap" --json --limit 10

# Pipe file paths to other tools
vim $(rfx query "TODO" --paths)

Interactive TUI mode — run rfx query with no pattern to launch live search with keyboard navigation.

Dependency analysis

rfx deps src/main.rs              # Show direct imports
rfx deps src/config.rs --reverse  # What imports this file
rfx deps src/api.rs --depth 3     # Transitive dependencies
rfx analyze --circular            # Find circular dependency chains
rfx analyze --hotspots            # Most-imported files
rfx analyze --unused              # Files with no incoming dependencies

Natural language search

rfx ask "Find all TODOs in Rust files"         # Translate to rfx query and run
rfx ask "How does authentication work?" --agentic  # Multi-step codebase reasoning
rfx ask                                        # Interactive chat mode

Requires an AI provider configured via rfx llm config (OpenAI, Anthropic, OpenRouter, or any OpenAI-compatible endpoint).

Other commands

rfx index                 # Build / update the search index
rfx index status          # Background indexing status
rfx watch                 # Auto-reindex on file changes
rfx stats                 # Index statistics
rfx pulse changelog       # Codebase change digest
rfx pulse wiki            # Per-module documentation
rfx pulse map             # Architecture diagram (Mermaid / D2)
rfx serve --port 7878     # Local HTTP API server

Run rfx <command> --help for full options.

Installation

NPM (recommended)

npm install -g reflex-search

Cargo

cargo install reflex-search

Setup note: run rfx commands from your project root directory. Add .reflex/ to your .gitignore to exclude the search index from version control.

Supported languages

Full symbol extraction (functions, classes, methods, types, etc.) for 15 languages:

Systems: Rust, C, C++, Zig
Backend: Python, Go, Java, C#, PHP, Ruby, Kotlin
Frontend: TypeScript, JavaScript, Vue, Svelte

Swift is temporarily disabled (tree-sitter-swift 0.7.x grammar incompatibility). rfx query --lang swift emits a warning; full-text search still works.

Full-text search works on all file types regardless of parser support.

Configuration

# .reflex/config.toml (project-level)
[index]
languages = []          # Empty = all supported languages
max_file_size = 10485760  # 10 MB

[search]
default_limit = 100

[performance]
parallel_threads = 0    # 0 = auto (80% of available cores)

For AI provider configuration (rfx ask, rfx pulse), run rfx llm config.

Architecture

Reflex uses a trigram-based inverted index with runtime symbol detection:

• Indexing: extracts 3-character trigrams from all files; stores full content in memory-mapped content.bin; no tree-sitter parsing at index time
• Full-text queries: intersect trigram posting lists → verify matches (instant)
• Symbol queries: trigrams narrow candidates → parse only matching files with tree-sitter

.reflex/
  meta.db          # SQLite: file metadata, stats, config
  trigrams.bin     # Inverted index (memory-mapped)
  content.bin      # Full file contents (memory-mapped)
  config.toml      # Index settings

Security

rfx serve binds to 127.0.0.1:7878 by default — loopback only, no authentication. Do not expose it to the network. See CLAUDE.md for the full threat model.

Contributing

cargo build --release   # Build
cargo test              # Test
rfx index               # Refresh index after code changes