Docs Case Studies

One query instead of
five greps.

cpk scan indexes your codebase with tree-sitter. Agents query symbols, imports, and dependents through one CLI — at the cost of a single grep, with the structural detail of reading every file. Plus a local task board for multi-agent coordination. 100% Open Source.

$ npm i -g codepakt
codepakt dashboard — kanban board with tasks assigned to claude and codex agents

The discovery tax

Most of an agent's "exploration" cost isn't reading the file it actually needs to change. It's the four to six grep cycles before that — finding what exists, where it lives, and what shape it has. cpk scan indexes that once, so agents replace those greps with a single structured query.

search

One query instead of five greps

cpk code symbols --name Broadcast returns every match across backend, frontend, and webhooks — with file paths, line numbers, signatures, and parent classes. No follow-up reads to figure out what's a class vs a method.

speed

Discovery, not verification

Signatures tell agents what exists. Bodies still need reading to verify behaviour. cpk code accelerates the first half of every session — orientation — at near-zero cost.

block

Atomic Task Pickup

When multiple agents coordinate on a shared board, BEGIN IMMEDIATE transactions guarantee exclusive pickup — no duplicate work.

Where the win scales
Familiar codebase marginal
Cross-module question ~30-40%
Unfamiliar codebase largest

The win isn't a flat percentage — it scales with how unfamiliar the code is. Agents landing in a fresh repo benefit most. Agents working in code they already know save less. cpk scan --install-hook keeps the index fresh so the value doesn't decay.

From zero to indexed in 30 seconds

One install, one scan, and your agents query a local index instead of burning tokens on exploration.

01

Install & Init

npm i -g codepakt
cd your-project && cpk init --name your-project

No Docker. No config. macOS + Linux.

02

Scan & Hook

cpk scan
cpk scan --install-hook

Tree-sitter parses TS/JS/Python/Go. Incremental updates on every git commit.

03

Agents Query

cpk code symbols --name X
cpk code dependents --file Y

~200 tokens per query. No server needed for code commands.

search

Code Intelligence

cpk scan parses your codebase with tree-sitter WASM and indexes symbols, imports, and dependents into a local SQLite file. cpk code symbols --name X returns every match — backend, frontend, webhooks — with file paths, line numbers, signatures, and parent classes in one structured payload. Works offline. No server needed.

layers

Atomic Task Pickup

BEGIN IMMEDIATE transactions prevent two agents from ever claiming the same task. Zero merge conflicts.

Git Hook Integration

cpk scan --install-hook keeps the index fresh on every commit. Incremental updates in milliseconds.

Live Dashboard

Kanban board with drag-and-drop at localhost:41920 for the task coordination half.

A real session

"There is no schedule option for WA campaigns. Use cpk for code intelligence."

One prompt to a Claude Code session on a real production codebase. Here's how the agent used cpk code to land the fix.

01

One query, full picture

$ cpk code symbols --name Broadcast

Returned ~30 symbols across backend, frontend, and webhooks in a single payload — file paths, line numbers, signatures, and parent classes. The agent identified scheduleBroadcast, processScheduledBroadcast, ScheduleBroadcastDialog, and BroadcastActions — without opening a single file.

02

Diagnosis from signatures alone

The signatures answered the structural question — "scheduling infra exists" — at zero file-read cost. The agent now knew the bug wasn't missing functionality. It was a UX problem: users couldn't reach the dialog.

03

Two targeted reads, not five speculative ones

The agent then opened only the two files needed to verify the UX gap:

  • broadcasts/page.tsx — the wizard + dialog wiring
  • csv-broadcast-wizard.tsx — confirmed the absence of a schedule path

Without the index: 4-6 grep+read cycles across apps/api/, apps/web/, hooks, and webhooks — each returning raw lines with no structural context.

"I'd estimate this exploration used ~30-40% fewer tokens than a grep/glob/read approach would have for the same conclusion. The win scales with codebase unfamiliarity — for a known module it'd matter less."

— Claude Code session on wapisnap.com, diagnosing missing schedule UX in a WhatsApp campaign tool

The Open Source Stack

smart_toy
Agents
terminal
cpk CLI
dns
Private Server
dashboard
Web UI

100% Open Source

MIT licensed. SQLite embedded. No Docker required. Install and go.

MIT LICENSE

Generated Coordination Files

cpk generate writes .codepakt/AGENTS.md and .codepakt/CLAUDE.md with the full command reference — so agents that drop into the project automatically know what to run.

# .codepakt/AGENTS.md — generated by `cpk generate`
## Code Intelligence (offline, no server needed)
cpk scan — full scan
cpk code symbols --name AuthService
cpk code dependents --file src/shared/types.ts
cpk code imports --file src/auth/guard.ts
cpk code summary
## Task Coordination (needs cpk server)
cpk task mine --agent <name>
cpk task pickup --agent <name>
cpk task done T-001 --agent <name> --notes "..."
## Active Agents
- claude (working on T-003)
- codex (idle)

Team

The people building codepakt.

SJ
Sameer Joshi
SM
SivaKumar Mettukuru

Ready to index?

Point codepakt at any repo and your agents stop burning tokens on discovery. MIT Licensed.

npm i -g codepakt
cd your-project
cpk init --name your-project
cpk scan
cpk scan --install-hook
cpk code summary