# Agentpack

> Local-first CLI and MCP server that keeps a repo-native task ledger so AI coding agents can resume tasks across sessions and clients.

Agentpack is an open-source CLI and local MCP server published as the npm package agentpack-cli. It keeps a small, reviewable task ledger in a `.agentpack/` folder inside a repository, so AI coding agents can record where a task stands and pick it up again in a later session. The latest version listed is 1.9.1.

## What It Is

Agentpack provides repo-native task continuity for AI coding agents. Connected agents record durable task state as they work: the goal, decisions, dead ends, verification evidence, and checkpoints. The next session loads that state and continues, whether it is the same agent after context compaction, a different MCP client, or a developer returning days later. The README describes it as complementary to agent memory: memory holds preferences and conventions, while Agentpack holds the state of a specific task.

## How the workflow runs

At session start the agent loads compact context: the current Task Passport, recent checkpoints, decisions, and reviewed conclusions about source files. While working, it records decisions, failed approaches, and verification evidence. Conclusions about files are stored with file hashes so later sessions can tell when a file has changed. At a natural stopping point the agent creates a checkpoint with status, next actions, and git state.

Each task has a Task Passport covering goal, status, constraints, write scope, next actions, and verification. Tasks can be started, parked, switched, and finalized. Resume output stays within a rough token budget.

## Client setup and task gate

After `agentpack init` in a repo, `agentpack install` configures a client: codex, claude, cursor, or claude-desktop. Installers generate project instructions such as `AGENTS.md`, `CLAUDE.md`, and Cursor rules. `agentpack doctor` verifies setup, and `agentpack resume` shows task state from the CLI. An optional task gate uses Claude Code, Codex, and Cursor hooks plus a pre-commit hook to warn when edits fall outside the active task, and it can be set to block. Codex, Claude Code, and Cursor installs also include an optional builder subagent.

## Security posture

The README states there are zero runtime dependencies, no telemetry, and no network calls during normal CLI or MCP operation. It also describes best-effort redaction of secret-looking values and releases published from GitHub Actions with npm provenance. Node.js 20 or later is required.

## Features
- Repo-local task ledger in .agentpack/
- Task Passport with goal, status, constraints, write scope, next actions, and verification
- Checkpoints with status, next actions, and git state
- Source conclusions stored with file hashes to flag changed files
- Local MCP server for agents
- Generated AGENTS.md, CLAUDE.md, and Cursor rules
- Optional task gate with warn or block mode via hooks and pre-commit
- Portable task bundles for handoff
- Token-budgeted resume output
- Best-effort redaction of secret-looking values
- Optional builder subagent
- CLI for inspection, doctor checks, and manual handoff

## Integrations
Claude Code, Codex, Cursor, Claude Desktop, MCP clients, Git hooks

## Platforms
WEB, API, CLI

## Pricing
Open Source

## Version
1.9.1

## Links
- Website: https://www.npmjs.com/package/agentpack-cli
- Documentation: https://github.com/ihorponom/agentpack/blob/main/docs/CLI.md
- Repository: https://github.com/ihorponom/agentpack
- EveryDev.ai: https://www.everydev.ai/tools/agentpack