Vexus, a newly released open-source tool for Claude Code users, tackles the inefficiency of agent-driven code navigation by pre-indexing repositories locally. Instead of letting the LLM burn tokens on iterative grep and read calls, Vexus injects a concise map of likely relevant code locations directly into the prompt. The project, hosted on GitHub by developer faique43, claims a 10% reduction in token costs across 288 live Claude Code sessions while maintaining identical pass rates on change requests.

The Token Tax of Blind Navigation

Standard agentic coding workflows suffer from what Vexus calls "the problem of full-text search." When an agent is asked how a feature works in an unfamiliar repo, it typically triggers a cascade of greps and file reads. This process pulls in dozens of irrelevant lines, inflating context windows and costs. Vexus intervenes by running a local hook on each prompt. It uses a local embedding model to find probable code locations and adds roughly 100 tokens of location hints to the agent's view. This allows the model to open the right lines first, bypassing the expensive discovery phase.

Local-First Architecture and Performance

The tool is built in Rust and relies on tree-sitter for parsing, supporting 18 languages including Python, TypeScript, and Rust. It uses ONNX Runtime for local embeddings, ensuring no code leaves the user's machine and no agent tokens are spent on indexing. Benchmarks on a 500-file Python repo show a cold index time of 17 seconds and search queries under 10 milliseconds. For larger repos like NestJS with over 2,000 files, the initial index takes about three minutes but remains efficient in background updates. The system employs hybrid retrieval, fusing vector KNN and BM25 keyword results with reciprocal rank fusion to balance semantic recall with exact-name precision.

Benchmark Results and Limitations

In controlled tests using Claude Sonnet 5.5, Vexus demonstrated a 21% cost reduction on short change requests and a 16% reduction in tool calls across all 288 sessions. However, the gains vary by task type; on real-world bugs from the NestJS codebase, savings dropped to 9%, and on tasks where the agent already knew the code location, savings were negligible. The tool also offers an optional MCP server for deeper exploration, though this adds overhead that can negate savings on simple change requests. Notably, the hook is currently exclusive to Claude Code, while other agents must rely on the MCP server, which yields less consistent results.

Key Takeaways

  • Token Efficiency: Vexus reduces Claude Code token costs by ~10% on average by pre-injecting code locations, avoiding repetitive grep/read loops.
  • Local Privacy: All indexing and embedding happen on-device using ONNX Runtime, with no code sent to external APIs.
  • Broad Language Support: The tool uses tree-sitter to support 18 languages, including Python, TypeScript, Rust, and Go.
  • Variable Savings: Cost reductions are highest on complex navigation tasks and lowest on simple queries where the agent already knows the file structure.
  • Platform Constraints: Full semantic search requires Apple Silicon macOS, Windows, or Linux with glibc 2.39+; Intel Macs and older Linux distros get a keyword-only fallback.

The Bottom Line

Vexus is a pragmatic answer to the high cost of agentic coding. By shifting the burden of code discovery from the LLM's context window to a local, pre-indexed map, it offers a tangible way to reduce token spend without sacrificing accuracy. While not a silver bullet for all agent tasks, its local-first architecture makes it a compelling tool for developers already deep in the Claude Code ecosystem who are tired of paying for the model to rediscover their own codebase.