jkr
Token-efficient AI dev workflows. Filter noisy command output, query code structurally instead of reading whole files, run agents in a sandbox — Apache-2.0, self-hosted, no telemetry.
jkr is two complementary halves of the same problem. Your AI coding assistant burns context on two things: (1) verbose command output (build / test / git / package managers), and (2) reading whole source files to find one function. jkr cuts both.
The two halves
1. CLI output filter (jkr <cmd>)
Wraps shell commands and strips dev-tool noise before it reaches the LLM context window. 60–90% token reduction on common operations.
jkr cargo test # only failures
jkr git status # filtered
jkr npm install # progress noise gone
jkr gain # how much you've saved
jkr watch # live dashboard (ratatui TUI)
Custom per-command filters live in ~/.jkr/filters/*.toml. Three rule types: suppress_prefix, suppress_regex, keep_regex. See Custom Filters below.
2. Code-intel MCP server (jkr-mcp)
Structured code-intelligence tools over MCP. Your agent reads outlines and signatures, not whole files. 1000×+ token reduction on "find the relevant code" tasks (measured: 5.7 MB of file content collapsed to 4.4 KB of indexed results across 4 realistic questions in this repo).
| tool | what it does |
|---|
jkr_index_build | One-shot persistent index of a repo (SQLite + tree-sitter, content-hash freshness) |
jkr_index_watch | Background file watcher — index stays fresh across edits |
jkr_outline_file | Symbols + line ranges for one file (no body) |
jkr_find_symbol | All definitions of a name across the repo |
jkr_signature | One-line declaration of a symbol |
jkr_read_smart | Ranked search by free-form question → top-K symbols with locations |
jkr_callers_of / jkr_callees_of | Cheap call-graph from the indexed refs table |
jkr_grep_summary | Regex grep with per-file aggregation + caps |
Supports 9 languages: Rust, Python, Go, TypeScript, JavaScript, Java, C, C++, Ruby. Set JKR_TOON=1 for compact TOON (Token-Oriented Object Notation) tabular output — ~15% additional savings on top of the structured queries.
Wire it into Claude Code with one command:
jkr mcp install # writes ./.mcp.json for the current repo
jkr mcp install --scope=user # or update ~/.claude.json so jkr is available in every project
jkr mcp install --print # dry-run: show the snippet without writing
The index builds on first query — no manual jkr_index_build step needed. (Optional: call jkr_index_watch once if you want the index to auto-refresh on file edits across long sessions.)
Tool descriptions in jkr-mcp are written to compete directly against native Read / Grep for the agent's attention — each tool's first line says which native tool it should replace and the byte-cost ratio. No CLAUDE.md hint required.
3. (Early) sandboxed agent runtime
jkr-agent + jkr-sandbox run agents with Landlock isolation. Bring your own provider (Anthropic, OpenAI, Ollama). Less mature than (1)/(2) — usable, not yet a headline feature.
What's different
- Plugin architecture — core is thin; filters and analytics are independent plugins on a shared bus (
cli.invoke routes to plugins that declare CLI subcommands).
- Encrypted vault — all plugin state lives in
~/.jkr/vault/ encrypted with age (XChaCha20-Poly1305); master key in ~/.jkr/vault/.jkr-vault.key (0600). One-time migration from legacy OS-keychain installs. Manage with jkr vault {status,init,unseal,seal,rotate,export,import,audit}.
- Noise analytics & suggestions —
jkr gain, jkr suggest, jkr watch read vault-backed analytics. Optional embeddings (--features embeddings) improve clustering during jkr suggest.
- Live dashboard —
jkr watch opens a ratatui TUI showing real-time token savings.
- Hooks —
jkr hook claude for Claude Code Bash hooks; jkr hook universal for the same JSON reply shape plus a top-level "command" field for other wrappers.
Install
Homebrew (macOS / Linux)
brew tap einyx/jkr
brew install jkr
Curl-to-bash
curl -fsSL https://raw.githubusercontent.com/einyx/jkr/main/install.sh | bash
Windows (x86_64)
Download jkr-x86_64-pc-windows-msvc.tar.gz from Releases. Extract jkr.exe using Git Bash, Windows 11 tar, or another tar; add the directory to PATH, then run jkr from PowerShell or CMD. After a manual install, jkr update pulls the matching jkr-<rust-target-triple>.tar.gz asset.
From source
Requires Rust 1.88+ (see rust-toolchain.toml).
git clone https://github.com/einyx/jkr
cd jkr
cargo build --release
cargo install --path crates/jkr --locked --force
