GitHub stars npm version MIT license

Code intelligence
for AI agents

Your AI agent wastes 70% of tokens reading irrelevant files. Sverklo indexes your codebase locally, so agents find the right code instantly — fewer tokens, faster results, better answers.

1. Add to Claude Code
$ claude mcp add sverklo -- npx sverklo .
2. Ask your agent anything
> "How does the auth middleware work?"
3. It searches & remembers
⚡ Found 4 results in 23ms — hybrid search across 1,200 files
Works with Claude Code Cursor & any MCP client

The problem

AI agents are blind without context

70%
of tokens wasted
AI agents read irrelevant files via grep, burning through your context window and rate limits.
10-30 min
lost per session
Developers rebuild context every session. Decisions, patterns, and preferences vanish after compaction.
<50ms
with Sverklo
Semantic search finds exactly the right code. Memories persist across sessions. Your agent stays smart.

Features

Search that understands your code

Hybrid retrieval with real ONNX embeddings, graph-based ranking, and session memory. Everything runs locally.

Hybrid Search

BM25 text matching combined with vector semantic search, PageRank graph ranking, and reciprocal rank fusion for precise results.

Real ONNX Embeddings

all-MiniLM-L6-v2 runs locally via ONNX runtime. No API keys, no network calls, no data leaving your machine.

10 Languages

TypeScript, JavaScript, Python, Go, Rust, Java, C, C++, Ruby, and PHP. Tree-sitter parsing for accurate symbol extraction.

Session Memory

Remember decisions, patterns, and context across sessions. Git-state linking, quality scoring, and automatic staleness detection.

Zero Config

Auto-indexes your codebase, respects .gitignore, incremental updates with file watcher. Just run it and search.

Token-Budgeted

Responses are trimmed to fit within LLM context windows. Your agent gets the most relevant code without wasting tokens.

TypeScript JavaScript Python Go Rust Java C C++ Ruby PHP

How it works

From code to answers in milliseconds

30 files indexed in 500ms. Search results in under 50ms. All local.

1

Watch

File watcher detects changes, respects .gitignore

2

Parse

Tree-sitter extracts symbols, types, references

3

Embed

ONNX model creates vector embeddings locally

4

Index

BM25 + vector index + PageRank graph built

5

Search

Hybrid retrieval with RRF fusion, ranked results

MCP Tools

10 tools your agent actually needs

Code search and session memory designed for how AI agents work.

Code Search
search
Hybrid code search across your entire codebase. Combines text matching, semantic similarity, and graph ranking.
overview
Get a high-level structural overview of the project: key modules, entry points, architecture patterns.
lookup
Jump directly to a symbol definition. Retrieve function signatures, class definitions, type declarations.
find_references
Find all usages of a symbol across the codebase. Understand impact before refactoring.
dependencies
Map import graphs and dependency chains. See what depends on what.
index_status
Check indexing progress, file counts, and embedding status for the current project.
Session Memory
remember
Save a decision, pattern, or context to persistent memory. Linked to current git state automatically.
recall
Retrieve relevant memories for the current context. Quality-scored and staleness-aware.
forget
Remove outdated or incorrect memories. Keep your memory store clean and accurate.
memories
List and browse all stored memories. Filter by topic, recency, or quality score.

Pricing

Open core, free forever

Full-featured code search and memory, MIT licensed. Pro and Team tiers are coming for advanced needs.

Available now

Free

$0

MIT licensed, forever

  • Hybrid code search (BM25 + vector + PageRank)
  • Real ONNX embeddings, fully local
  • 10 language support
  • Session memory with git linking
  • 10 MCP tools
  • Incremental indexing & file watcher
  • Token-budgeted responses
Get started

Pro

TBD / month

Coming soon

  • Everything in Free
  • Better embedding models
  • Cross-project pattern detection
  • Search analytics & insights
  • Priority support
Coming soon

Team

TBD / seat / month

Coming soon

  • Everything in Pro
  • Shared team memory
  • On-premise deployment
  • Agent coordination
  • Admin & access controls
Coming soon

Quick Start

Up and running in 60 seconds

1

Download the embedding model (one-time, ~90MB)

npx sverklo setup
2

Start the MCP server in your project directory

npx sverklo .
3

Add to Claude Code (or any MCP client)

claude mcp add sverklo -- npx sverklo .
4

Start coding. Your agent now has deep codebase search and persistent memory.

Copied to clipboard