Relith

Relith is a local-first context engine for AI-assisted coding. It indexes your codebases and exposes them through a unified MCP interface - one index, any AI.

Instead of every AI tool building its own isolated context, Relith is a shared intelligence layer: Cursor, Claude Code, OpenCode, and any MCP client query it for code search, symbol lookup, reference tracking, dependency graphs - and now, git history.

Why Relith?

• One index, every AI. Every AI tool asks the same questions of your codebase. Relith builds the context once and serves every MCP client from the same index - no per-tool re-indexing, no duplicated state.
• Local-first by default. One self-contained download, no runtime or cloud account required. Docker is optional - a container image is included for those who want it. Your code never leaves your machine.
• Beyond grep. Search only finds strings. Relith also answers where is this used, what calls this, how do these files relate, and when, why, and who changed this line - via symbols, references, a dependency graph, and git-aware tools.
• Zero glue code. Speak the Model Context Protocol over stdio and your tools show up in any MCP-compatible agent automatically.

How it compares

	ripgrep	ast-grep	Sourcegraph	Relith
Search	Regex over files	Structural (AST) patterns	Regex + keyword + semantic (hosted)	Regex/FTS over a shared index
Multi-repo	Per invocation	Per directory	Yes (hosted)	Yes - one index
Symbols & references	No	Partial (structural)	Yes (SCIP)	Yes, 17 languages
Callers/callees & graph	No	No	Yes (code navigation)	Yes (knowledge graph)
Git-aware context	No	No	Yes (hosted)	Yes (commits, history, blame, diffs)
MCP-native	No	No	Yes (Enterprise instance)	Yes, local stdio
Data locality	Local	Local	SaaS or self-host	Local
Setup	One binary	One binary	Instance + account	Tarball + one-liner install

Features

• MCP-native - 21 tools for AI assistants: search, symbols, references, definitions, callers/callees, file outline, dependency tracing, graph queries, architecture overview, commits, blame, and diffs
• Git-aware context - ask when / why / who: recent commits, per-file history (follows renames), per-line blame, and full patches between any two refs
• Cross-file reasoning - one trace_context bundle combining FTS search + symbol matches + references + graph neighbors
• Knowledge graph - typed dependency graph (import edges for Go/JS/TS/Python/Rust, reference co-occurrence for all), visualized with an interactive D3.js force-directed graph
• Symbol & reference extraction - functions, types, methods, interfaces, enums, macros across 17 languages
• Multi-repo - index unlimited repos, search across all at once
• Self-healing watcher - auto-reindexes changed files via fsnotify
• Terminal UI - Bubble Tea progress bars, spinners, and a server dashboard
• REST API - HTTP server for scripts, CI pipelines, and programmatic access
• Local-first, zero runtime - static Go binaries in one download, no npm/pip/uv, no runtime, Docker optional, your code never leaves your machine

Performance

Linux kernel, 94,989 files, 1.7M chunks:

Phase	Time
Walk + index	14m 40s
Graph build	1m 8s
Total	15m 48s

Reproduce it on your machine with bench/index-kernel.sh - see docs/benchmarks.md.

Quick Start

git clone https://github.com/cryskram/relith.git && cd relith
make build-all

# ...or install the latest release:
curl -fsSL https://raw.githubusercontent.com/cryskram/relith/main/install.sh | sh
# On Windows, run the same command from Git Bash / MSYS2.

# ...and to uninstall (add --purge to also delete the index database and config):
curl -fsSL https://raw.githubusercontent.com/cryskram/relith/main/uninstall.sh | sh

./bin/relith repo add /path/to/your/project
./bin/relith index
./bin/relith search "your query"

# Wire your AI agent (OpenCode, Cursor, Claude Code)
./bin/relith install

# Daemon: REST API + graph UI + file watcher
./bin/relith serve

macOS: If Gatekeeper blocks the downloaded binary, run xattr -d com.apple.quarantine /path/to/binary.

CLI

Interactive commands (index, remove, serve) render a Bubble Tea TUI automatically; otherwise they print plain text.

Command	Description
relith repo add <path>	Register a repository for indexing
relith repo list	List all indexed repositories
relith repo remove <id-or-name>	Remove a repository and all its data
relith index [path]	Index a repo (or all pending)
relith search <query>	Full-text search across all indexed code
relith status	Show indexing status with file/chunk counts
relith serve	Start the daemon (REST API + graph UI + file watcher)
relith install	Auto-configure MCP for OpenCode, Cursor, Claude Code
relith uninstall	Remove relith MCP configuration from agents
relith db vacuum	Reclaim unused database space
relith version	Print the version
relith search <query> --limit=N	Limit search results

MCP Server

Relith speaks the Model Context Protocol over stdio. Point any MCP-compatible assistant at relithmcp, and the tools show up automatically - no glue code.

relith install                 # auto-detect installed agents
relith install --agent=cursor  # or target a specific one
relith install --agent=code    # Claude Code

Git-aware tools

On any repo that's a git worktree, the AI can also ask:

Tool	Answer
get_recent_commits	What's been done recently (hash, author, date, message)
get_file_history	What has this file been through
get_blame	Who owns each line of this file
get_diff	Exactly what a commit / PR changed (stat + patch)

These shell out to your system git, so they always match the repo's real state.

Full tool list

Category	Tools
Search	search_code, find_symbol, find_references, find_callers, find_callees
Files	get_file_content, get_file_outline, get_file_tree, get_repo_summary
Graph	get_related_files, list_hub_files, query_graph, get_architecture, trace_dependency
Git	get_recent_commits, get_file_history, get_blame, get_diff
Reasoning	trace_context, get_symbol_definition, list_repositories

Manual setup

Agent	Where
Cursor	Settings → MCP Servers → Add: command /path/to/relithmcp
Claude Code	~/.config/claude/mcp.json → {"mcpServers": {"relith": {"command": "/path/to/relithmcp"}}}

REST API

Via the daemon:

curl -s 127.0.0.1:9876/v1/health
curl -s "127.0.0.1:9876/v1/search?q=sqlite"
curl -s 127.0.0.1:9876/v1/graph          # interactive graph (browser)
curl -s 127.0.0.1:9876/v1/graph?repo=my-repo  # graph data (JSON)

Configuration

~/.config/relith/relith.yaml (or %LOCALAPPDATA%\Relith\relith.yaml). Any key overrides with a RELITH_ env var (RELITH_INDEXER_CONCURRENCY=8).

core:
  data_dir: ~/.local/share/relith
daemon: { tcp_host: 127.0.0.1, tcp_port: 9876 }
mcp:    { enabled: true, transport: stdio }
indexer: { concurrency: 4, max_file_size: 10485760 }
watcher: { enabled: true, debounce: 1s }
search: { max_results: 100, path_boosting: true }

How it works

Three binaries share one SQLite (FTS5 + WAL) database:

Binary	Role
relith	CLI + TUI
relithd	REST API + graph UI + file watcher
relithmcp	MCP server over stdio

Walk → chunk → extract symbols & refs → build dependency graph → serve via MCP, HTTP, or TUI.

See ARCHITECTURE.md for the full picture.

License

MIT - see LICENSE.

Contributing

See CONTRIBUTING.md for the dev setup and contribution workflow. Security issues: see SECURITY.md.