Docs Case Studies

Getting Started

codepakt is the agent context layer for AI coding agents. Two things, one CLI:

  1. Code intelligence. Agents waste 6,000-40,000 tokens per session exploring codebases via grep/glob/read cycles. cpk scan parses your codebase with tree-sitter and stores symbols and imports in SQLite. cpk code symbols/imports/dependents/summary answers “where is X?” and “who uses Y?” in ~200 tokens each.
  2. Task coordination. A shared board that agents self-serve from. Atomic pickup, dependency resolution, and a local kanban dashboard. No duplicate work, no merge conflicts, no ceremony.

Everything ships in one npm package (codepakt) with no MCP server, no LLM keys, and a ~250 token cost per CLI call.

Who it’s for

  • Developers running AI agents on large codebases who keep burning tokens on grep and ls
  • Teams running 2+ agents in parallel where the agents need to divide work without stepping on each other
  • Anyone who wants a local, auditable, offline-capable agent context layer that works from any terminal

Install

npm i -g codepakt

Requires Node 20+. macOS and Linux only.

Code intelligence in 30 seconds

From the root of any project:

cd your-project
cpk init --name your-project        # Register the project
cpk scan                             # Parse everything with tree-sitter WASM
cpk code summary                     # Overview: files, symbols, languages
cpk code symbols --name AuthService  # Find the symbol you're looking for
cpk code dependents --file src/shared/types.ts  # Who imports this?

cpk scan and cpk code work offline. They read/write .codepakt/data.db directly. No daemon required. Run cpk scan --install-hook to auto-update the index on every git commit.

Task coordination in 30 seconds

Task commands need the server daemon for atomic pickup across concurrent agents:

cpk server start                        # Start the coordination daemon
cpk task add --title "Build auth API" --priority P0
cpk task pickup --agent dev             # Atomic claim
cpk task done T-001 --agent dev --notes "Implemented JWT auth"
open http://localhost:41920             # Kanban dashboard

That’s the full loop. No agent registration. No config. The server tracks state. The agents do the work. The dashboard shows you what’s happening.

What makes it different

~200-250 tokens per CLI call. Compare to 5,000-8,000 tokens for an MCP server that injects full schemas on every call. At scale with multiple agents and long sessions, the gap compounds.

Offline-first code intelligence. cpk scan and cpk code read the local SQLite file directly — no HTTP layer, no daemon. Agents can query the index even when the server isn’t running.

Dumb server, smart agents. The server has zero AI. No LLM keys, no opinions about your codebase. Agents do the planning, scanning, and reasoning. The server stores state and enforces concurrency for shared operations (task pickup).

Works with any agent that can run bash. Claude Code, Codex, Cursor, Copilot, shell scripts — if it can call cpk, it can use the index and participate in the board.

Atomic task pickup. Two agents can’t claim the same task. The daemon uses BEGIN IMMEDIATE transactions for race-free pickup.

Next steps

Next
Installation
arrow_forward