grepika

Token-efficient MCP server for code search.

LLMs burn context tokens on every search call. grepika indexes your codebase and returns ranked, compact results — so the model spends tokens reasoning instead of reading raw grep output. It combines three backends (FTS5, parallel grep, and trigram indexing) and merges their scores for high-quality results.

Why grepika?

grep is a great tool, but it wasn't designed for LLM workflows. It returns unranked file lists, and the model has to make multiple calls to piece together context. grepika gives the model structured tools for the things it actually needs to do:

Benchmarks

Criterion benchmarks against ripgrep (Claude Code's Grep backend) on the grepika codebase, 9 queries across all intent categories:

~80% fewer bytes on aggregate vs ripgrep content mode, with ranked results and snippets. Savings are largest on high-match queries (e.g. fn saves 94%); natural language queries where ripgrep finds few literal matches can be larger with grepika. See full analysis.

Token Efficiency

Compared to ripgrep content output (matching lines), indexed search returns fewer tokens per query. The bigger win is search quality — ranking, NLP queries, and reference classification reduce follow-up reads.

Criterion benchmarks on the grepika codebase (~25 Rust files), 9 queries across all intent categories:

Query             │  grepika │  ripgrep (content) │ Savings
──────────────────┼──────────┼────────────────────┼────────
SearchService     │  3,375 B │         8,610 B    │  ~61%
Score             │  2,789 B │         5,574 B    │  ~50%
Database          │  2,156 B │        10,174 B    │  ~79%
fn                │  1,832 B │        31,469 B    │  ~94%
use               │  1,963 B │        23,308 B    │  ~92%
search service    │  3,797 B │         1,632 B    │ -133%
error handling    │  2,666 B │           986 B    │ -170%
fn\s+\w+          │    385 B │        29,501 B    │  ~99%
impl.*for         │  3,214 B │         2,480 B    │  -30%

Savings are largest on high-match queries where ripgrep returns many unranked lines. Natural language queries (e.g. "error handling") route to FTS5 concept search in grepika but match few literals in ripgrep, making grepika's output larger.

Claude Code lazy-loads MCP tools on demand, so grepika's 11 tool schemas are not loaded all at once. Loaded schemas are prompt-cached after the first call (~90% discount on subsequent turns). In practice, schema overhead is minimal.

See docs/token-efficiency-analysis.md for the full comparison including Grep file-list mode and workflow analysis.

How it works

• Three search backends (FTS5 + grep + trigram) with weighted score merging
• BM25 ranking with tuned column weights
• Query intent detection — classifies regex vs natural language vs exact symbol
• 190 tests, zero clippy warnings, Criterion benchmarks

MCP Server Setup

By default, grepika runs in global mode — the server starts without --root, and the LLM calls add_workspace with its working directory automatically.

/plugin marketplace add agentika-labs/agentika-plugin-marketplace
/plugin install grepika@agentika-labs-agentika-plugin-marketplace

# For all your projects (user-level)
claude mcp add -s user grepika -- npx -y @agentika/grepika --mcp

# For this project only (shared with team via .mcp.json)
claude mcp add -s project grepika -- npx -y @agentika/grepika --mcp

## Code Search

Prefer grepika MCP tools over built-in Grep/Glob for code search:
- `mcp__grepika__index` - Build/update search index (run first!)
- `mcp__grepika__search` - Pattern/regex search (replaces Grep)
- `mcp__grepika__toc` - Directory tree (replaces Glob patterns)
- `mcp__grepika__outline` - File structure extraction
- `mcp__grepika__refs` - Symbol references

These provide ranked results with FTS5+trigram indexing for better search quality.

{
  "permissions": {
    "allow": [
      "mcp__grepika__*"
    ]
  }
}

{
  "mcpServers": {
    "grepika": {
      "command": "npx",
      "args": ["-y", "@agentika/grepika", "--mcp"]
    }
  }
}

{
  "mcp": {
    "grepika": {
      "type": "local",
      "command": ["npx", "-y", "@agentika/grepika", "--mcp"]
    }
  }
}

{
  "mcpServers": {
    "grepika": {
      "command": "npx",
      "args": ["-y", "@agentika/grepika", "--mcp"]
    }
  }
}

claude mcp add -s user grepika -- npx -y @agentika/grepika --mcp --root /path/to/project

{
  "mcpServers": {
    "grepika": {
      "command": "npx",
      "args": ["-y", "@agentika/grepika", "--mcp", "--root", "/path/to/project"]
    }
  }
}

Tip: Add "--db", "/path/to/index.db" to args to control where the index is stored.

CLI Setup

npm

npx -y @agentika/grepika <command>

Shell script (macOS Apple Silicon)

curl -fsSL https://raw.githubusercontent.com/agentika-labs/grepika/main/install.sh | bash

For other platforms, download the binary from GitHub Releases.

Usage

# Index a codebase
grepika index --root /path/to/project

# Search (modes: combined, fts, grep)
grepika search "authentication" --root /path/to/project -l 20 -m combined

# Get file content with line range
grepika get <path> -s 1 -e 100

# View index statistics
grepika stats

# Extract file structure (functions, classes, structs)
grepika outline <path>

# Directory tree
grepika toc --root /path/to/project -d 3

# Surrounding context for a line
grepika context <path> -l 42 -c 10

# Find all references to a symbol
grepika refs <symbol>

# Compare two files
grepika diff <file1> <file2>

# Generate shell completions
grepika completions <shell>

# Run as MCP server (global mode — LLM calls add_workspace)
grepika --mcp

# Run as MCP server (single workspace mode)
grepika --mcp --root /path/to/project

Available Tools

Configuration

Index Location

By default, the index is stored in a global cache directory, not in the project:

The <hash> is derived from the absolute path to --root, ensuring each project gets its own index without polluting the project directory.

Use --db to specify a custom location:

grepika --mcp --root /path/to/project --db /custom/path/index.db

Other Settings

• Max file size: 1MB (files larger than this are skipped during indexing)
• Gitignore: Patterns in .gitignore are respected during indexing
• Logging: All logs go to stderr (stdout is reserved for JSON-RPC in MCP mode)

Contributing

See CONTRIBUTING.md for build, test, benchmark, and profiling instructions.

Star History

Built by

Agentika — we help teams configure and adopt AI tools. If you need help setting up grepika or other AI dev tools for your team, get in touch.

License

MIT