RLS is local code search for AI agents with native MCP support. Single exe, CLI for humans plus MCP for Claude Code / Codex, runs 100% on your machine and cuts code-reading tokens by 91.3% on sample tasks. Free download with login at rls.swancat.com.

Token figure is a reference magnitude on fixed sample tasks, not a promise. Method in docs/BENCHMARKS.md, walkthrough in docs/TRIAL-REPORT.md.

91.3% fewer tokens for deep code inspection on fixed sample tasks (search+reads -84.5%, overview -85.9%, deep inspect -95.0%).

Local only, no code upload, listens on 127.0.0.1 by default, no telemetry.

3 calls, ~80ms from unknown symbol to definition + callers in the real trial (search → context → ref).

What does RLS do?

RLS is a local code search MCP and grep alternative for AI: instead of the agent doing grep + full-file reads (refilling context with redundant code), it locates symbols and traces call paths itself — locate symbol/file, inspect structure/definition, trace calls on demand. Returns file + symbol + line first; AI deep-reads only what matters.

How many tokens does RLS save?

About -91.3% total on the fixed tasks measured (Windows x64): the saving comes from reading fewer files because locating is precise, plus --compact and bounded trace compressing payloads. Your repo will differ — see the reproducible method in the repo and bring your own numbers to Issues.

Speed side: warm query p95 ~20ms (50 Rust files), exact search ~3.74ms, trace 4K bounded ~30ms; 500 files warm ~27ms, 5,000 files warm ~91ms.

Does RLS upload my code to the cloud?

No. RLS runs 100% on your machine: code and index never leave it, listens on 127.0.0.1 by default, no telemetry, no auto-update checks, no outbound requests. Single binary; verify with packet capture or firewall.

Is RLS open source?

No. Binary + docs only (docs-only repo, no source, no binaries in the repo). Downloading means you accept the LICENSE. Free trial, Windows x64 first.

How do I use RLS with Claude Code?

Quickstart (download):

  1. Get rls-vX.Y.Z-windows-x64.zip at rls.swancat.com (free, login with your Swancat account), unzip to get rls.exe, verify SHA256 on the download page.
  2. .\rls.exe install then .\rls.exe service start, endpoint http://127.0.0.1:8765/.
  3. The installer tries to write the user-level MCP entry; if it fails, add it manually. Check with claude mcp get rls / codex mcp get rls --json. Humans use the same index via CLI (search / outline / context / trace / handover).

What languages does RLS support?

Structured: Rust, Python, JavaScript, TypeScript, Go, Java, C, C++. Others remain text-searchable. macOS/Linux: core is cross-platform, Windows x64 ships first — vote in Issues.

Pair with HiRouter

Pair with HiRouter BYOK gateway to cut tokens at both the input and gateway layers, for maximum cost savings.

Brand: RLS / SwanCat. Updated 2026-09-14.