xlr8

GitHub in a box: instant local web viewer for any repo. One command, one ~5 MB binary. Commits, compare, search and PRs, served from your own machine. Runs on Windows, Linux, and even a phone (Termux).

xlr8 owner/repo

• Repeat views in ~0.1 ms. Results are cached in memory and on disk, so they survive restarts.
• First compare 1.7–4.8× faster than git diff on 7 of 8 benchmark ranges (16 threads, Windows).
• Git-accurate. Exact totals on synthetic ranges, within 0.16% on real repos, identical rename pairs.

_{tokio: 4,783 commits, browsed locally}

Install

Download a prebuilt binary from Releases (Windows, Linux x86_64), unzip / untar, and run xlr8 owner/repo.

Or build from source:

git clone https://github.com/Banshal-Yadav/xlr8.git
cd xlr8
cargo build --release

Binary: target/release/xlr8 (~5 MB). Requires the Rust toolchain (rustup). Other platforms (macOS, ARM/Termux): build from source.

Usage

xlr8 owner/repo --port 7782     # open one repo
xlr8 --serve --port 7782        # serve all previously opened repos

Opens http://127.0.0.1:7782. The repo is mirrored to ~/.cache/xlr8/mirrors/ and synced on request. Flags: --port <p> (default 7782) · --serve (multi-repo) · --no-open (don't launch browser).

What it does

• Commits: paginate, per-commit +/−, activity strip, author rank
• Compare: rename detection, edit-script diffs, windowed rendering
• Search: fuzzy path matching, in-repo grep, blob viewer
• Pulls: browse open PRs
• Perf bar: live latency sparkline, cache meter, RAM/CPU

Performance

Windows 11 · Ryzen 7 7735HS (8C/16T) · NVMe. Medians. Baseline is git diff --numstat -M on the same packed mirror (includes process spawn); xlr8 times are full HTTP round-trips with JSON. Cold = fresh process, all caches wiped (5 runs). Warm = 100 repeats over a keep-alive connection.

• Repeat views are effectively free: ~0.1 ms from memory; compare results are also cached on disk by commit OID.
• First views beat git on 7 of 8 ranges with the default thread count. Single-threaded, cold is roughly on par with git, so most of the cold margin comes from parallel diffing.
• Known weak spots: very small ranges and single-threaded deep delta chains.

Full tables, per-run numbers and method: docs/BENCHMARKS.md.

Also runs on a phone

xlr8 runs on Android via Termux (tested on a Snapdragon 695 / Poco X4 Pro 5G, aarch64). Early numbers from that device are in docs/benchmark.md; they use an older harness and PRoot, so they are not comparable to the table above.

HTTP API

• GET /api/{owner}/{repo}/refs — branches, tags
• GET /api/{owner}/{repo}/commits?limit=N&ref=REF&stats=1&count=1 — commit list
• GET /api/{owner}/{repo}/diff?base=REF&head=REF — compare summary (+/− per file)
• GET /api/{owner}/{repo}/search?q=... — path + grep search
• GET /api/{owner}/{repo}/file?base=...&head=...&path=... — file diff/hunks
• GET /api/repos — list mirrored repos
• GET /api/stats — runtime stats (RSS, CPU, cache, uptime)

Built with

gix · axum · tokio · imara-diff · vanilla HTML/JS

License

MIT © 2026 Banshal-Yadav