Ever watched your coding agent grep blind through a repo, open three files it didn't need, and burn fifteen thousand tokens just to answer "who calls this function"? That's the problem dekko exists to fix. dekko is a fast, offline, dependency-free static code map generator and codebase indexer for LLM coding agents. It scans a repository with tree-sitter (no model tokens spent parsing) and writes:
MAP.md— a human-readable map: a per-directory overview, an embedded architecture diagram, load-bearing/orchestrator rankings, then every file's functions/methods with signatures, doc lines, and who calls and is called by whom.map.json— the same graph in machine-readable form.
On top of the map, dekko gives an agent a token-cheap way to answer
questions like "what does this file contain," "who calls this function,"
and "what do I need to safely change this" — without reading whole files.
It ships as a CLI, a Claude Code /map plugin + MCP server
(Model Context Protocol), and
works with Cline too.
The result: 10x-300x fewer tokens than a plain Read/Grep workflow on the everyday tasks (orientation, outlining a large file, a symbol's callers), measured across 7 real, unmodified open-source repos on dekko 0.43.77. The floor, a small grep-friendly local symbol, is about 2.5x. The full breakdown is right below.
Why dekko?
Most agent workflows gather context by reading whole files or grepping
across a repo — expensive, and it throws away structure (who calls
what, what a function's fan-in/fan-out looks like). dekko instead
parses the repo once into a call graph and answers targeted questions
against it. Measured across 7 real, unmodified open-source repos
(Go, TypeScript, Java, Rust, Python/C++ — up to 14k files, 172k
symbols) on dekko 0.43.77, dekko's structured queries used 10x–300x
fewer tokens than the equivalent Read/Grep workflow for the same
task (repo orientation, outlining a large file, tracing a symbol's
callers).
dekko's cost stays roughly flat per query while Read/Grep scales
with file/repo size, so the ratio grows with scale. It's fast in
wall-clock terms too: mapping dekko's own ~4,300-symbol codebase from a
cold cache takes about 5 seconds on all cores, an incremental remap
after an edit is well under a second, and queries against the resulting
map return instantly. The win isn't universal — small, self-contained
files and already-grep-friendly local symbols see little benefit (the
floor measured was about 2.5x). See
benchmarks/real-world-repos/
for the full per-task breakdown, methodology, the original 2026-08
study, and the correctness caveats it raised and how they were closed.
Compared to tag-index tools like ctags/gtags, dekko resolves actual
call edges (not just definitions), ranks files by load-bearing-ness,
and speaks directly to agents over MCP or the CLI — no editor plugin
required.
Install
uv tool install dekko # or: pip install dekko / pipx install dekko
dekko --claude-install # add the /map command + MCP server to Claude Code, then restart
Extras (dekko[all] for 43 more languages, dekko[search] for
embedding search), installing from a local clone, and uninstalling are
in docs/install.md.
Quick start
cd my-project
dekko map # writes .dekko/MAP.md + .dekko/map.json
dekko summary # ~40-line digest: dirs, hotspots, entry points
.dekko/ is git-ignored by default; the map regenerates on demand, so
you rarely need to run dekko map again by hand. If your repo has
languages outside the default Tier-1 set (Python, C, C++, JS/TS, Go,
Java, Kotlin, Rust), dekko map will say so per file; install dekko[all] for
43 more languages (see Install) and re-run. The
list says how well each one is read.
Documentation
- docs/install.md — installation, extras, local clone, uninstall
- docs/cli.md — every CLI command, symbol targets, excluding files, notes, daemon mode, language support
- docs/claude-code.md — the
/mapplugin, push hooks, Claude Code skills, the MCP server, and Cline
Learn more
- CHANGELOG.md — per-version history
- CONTRIBUTING.md — dev setup, testing, releasing
- benchmarks/ — token-efficiency measurements, including a 7-repo real-world comparison against a plain Read/Grep workflow