ktlsp

ktlsp

A fast, lightweight language server for Kotlin and Java, written in Rust.

ktlsp focuses on the editor loop: quick startup, low overhead, useful navigation, and conservative answers. It uses tree-sitter and an in-memory index instead of keeping a compiler or JVM on the request path.

Features

Kotlin and Java support includes:

• definitions, references, implementations, and type definitions
• completion, hover, signature help, highlights, and symbols
• rename, auto-import, and unused-import actions
• semantic tokens, inlay hints, folding, and selection ranges
• call and type hierarchies
• navigation into dependency, JDK, and Android SDK sources
• fast, compiler-free diagnostics

Results are intentionally conservative: when the index cannot prove an answer, ktlsp may return no result instead of guessing. Kotlin type inference, overload resolution, and Gradle modeling are not yet compiler-complete.

Install

Download a binary from GitHub Releases, or install from source:

cargo install --git https://github.com/pepegar/ktlsp ktlsp

With Nix:

nix run github:pepegar/ktlsp

Building requires a Rust toolchain and a C compiler for the tree-sitter grammars.

Editor setup

ktlsp communicates over standard input and output. Configure your editor to start ktlsp for Kotlin and Java files.

Neovim example:

vim.lsp.start({
  name = "ktlsp",
  cmd = { "ktlsp" },
  root_dir = vim.fs.root(0, {
    "settings.gradle.kts",
    "settings.gradle",
    "build.gradle.kts",
    "build.gradle",
    ".git",
  }),
})

Logs are written to standard error. Set RUST_LOG=ktlsp=debug for debug logging.

Claude Code

The repository is also a Claude Code plugin marketplace. The ktlsp-lsp plugin connects Claude Code's built-in LSP tools to ktlsp for Kotlin, Kotlin Script, and Java files.

Install ktlsp first using one of the methods above and verify the binary is on the PATH used to start Claude Code:

command -v ktlsp

Then add the marketplace and install the plugin from Claude Code:

/plugin marketplace add pepegar/ktlsp
/plugin install ktlsp-lsp@ktlsp
/reload-plugins

Claude Code LSP plugins configure a connection to an installed language server; they do not install the server binary itself. If another Kotlin or Java LSP plugin is enabled, disable it to avoid an extension conflict. See the plugin README for CLI installation, troubleshooting, local validation, and community-marketplace publishing notes.

Configuration

ktlsp reads Gradle version catalogs and resolves source archives from local Gradle and Maven caches, Maven Central, installed JDKs, and Android SDKs. Extracted sources and symbol caches are stored in the platform cache directory.

Useful environment variables:

• KTLSP_CACHE_DIR — override writable cache storage
• KTLSP_JDK_SRC — point to a JDK src.zip
• KTLSP_ANDROID_SOURCES — point to an Android sources directory or archive
• KTLSP_ANDROID_SOURCES_DOWNLOAD=0 — disable Android source downloads

Formatting is opt-in through initializationOptions.formatting and delegates to an external formatter such as ktfmt.

Benchmarks

A July 2026 protocol-level comparison used Pinterest ktlint (multi-module JVM) and Square Okio (Kotlin Multiplatform), with one cold run and medians of three warm process restarts.

Workload	Metric	ktlsp	JetBrains Kotlin LSP	Relative result
ktlint, cold	First verified cross-file definition	0.51 s	11.12 s	ktlsp 21.7x sooner
ktlint, warm	First verified cross-file definition	0.51 s	1.98 s	ktlsp 3.9x sooner
ktlint, warm	Completion latency	0.8 ms	182.2 ms	ktlsp 228x faster
ktlint, warm	References latency	5.9 ms	58.3 ms	ktlsp 9.9x faster
ktlint, warm	Rename edit computation	5.4 ms	2,903.6 ms	ktlsp 538x faster
ktlint, warm	Post-index/import RSS	1,390.9 MiB	1,173.9 MiB	ktlsp uses 1.18x the RSS
Okio KMP, cold	First verified cross-file definition	0.51 s	No result within 60 s	ktlsp >116x sooner within timeout

ktlsp becomes useful much earlier and keeps foreground requests fast, but its fully loaded dependency/JDK index is not always smaller than JetBrains' persisted model. JetBrains also provides deeper compiler-backed diagnostics, quick fixes, formatting, and project import, so latency is not a semantic-equivalence score.

See BENCHMARK.md for exact revisions, host details, cold/warm cache rules, feature results, memory methodology, reproduction steps, and limitations.

Development

The workspace contains the shared semantic engine (ktcore), language server (ktlsp), and static-analysis CLI (ktcheck).

cargo test --workspace
cargo fmt --all -- --check
dev/ktlsp-harness.sh basic