Agents Pack — shared skills and portable project memory for coding agents

5 min read Original article ↗

Open source · Local-first CLI

Teach every coding agent how you work.

Agents Pack maintains one set of instructions, skills, and subagents, then renders them in each agent's native format. It also gives Claude Code, Codex, and Cursor portable, repository-owned project memory — without an account, a server, or anything overwritten behind your back.

Install — macOS and Linux

curl -fsSL https://farfarawaylabs.github.io/agentspackai/install.sh | sh

Works with
Claude Code · Codex · Cursor

In the core pack
42 skills · 7 subagents · 1 instruction set

Current pack
0.33.0

The problem

One capability. Three formats. Three copies to keep in sync.

Claude Code, Codex, and Cursor each expect instructions, skills, and subagents in different places, in different formats, with different discovery rules. Maintain them by hand and you write the same capability three times — then find out it drifted when an agent quietly stops following it.

You maintain

One canonical source

  • Core instructions
  • Skills
  • Subagents

Official pack content, plus anything you author yourself.

Agents Pack renders

  • Claude Code

    • .claude/rules/agents-pack/
    • .claude/skills/
    • .claude/agents/

  • Codex

    • AGENTS.md (managed block)
    • .agents/skills/
    • .codex/agents/

  • Cursor

    • .cursor/rules/agents-pack/
    • .cursor/skills/
    • .cursor/agents/

Repository-scope paths. Agents Pack records a hash for every file it writes, so a hand edit is reported as drift instead of being silently replaced.

Portable project memory

Project memory that follows you from agent to agent.

When Claude Code saves a verified project fact, Codex or Cursor can find it in a later session. Durable knowledge stays as readable Markdown inside the repository instead of being trapped in one provider's memory store.

  • Claude Code
  • Codex
  • Cursor

Inside your repository.agents-pack/memory/

Plain Markdown

MEMORY.mdCompact index of active shared memory

local/Git-ignored

Personal preferences, machine paths, and checkout-specific details.

No memory CLI · No database · No background process

  1. Cross-agent recall

    Core instructions prompt each agent to recall relevant memory during orientation and save verified, durable learning at natural checkpoints.

  2. Repository-owned

    Memory stays in ordinary Markdown that you can search, review, edit, and merge with the filesystem and Git tools you already use.

  3. Shared or local

    Team knowledge is Git-trackable by default. User preferences, machine paths, and checkout details stay in an ignored local scope.

See how portable memory works
How it works

Three commands, then it is out of your way.

Agents Pack is a filesystem tool. It writes native files, records what it owns, and gets out of the loop.

  1. Install the CLI

    One standalone executable with its checksum verified, placed in ~/.local/bin without sudo. No Bun, Node.js, or npm required.

    curl -fsSL https://farfarawaylabs.github.io/agentspackai/install.sh | sh

  2. Initialize your project

    Choose repository or global scope, the agents you use, and the components you want. Agents Pack shows the exact plan before it writes anything.

  3. Just work

    Each agent discovers the rendered files through its own native mechanism. Nothing to wire up, no extension to install, no process to keep running.

The core pack

42 skills and 7 subagents, maintained for you.

Skills are loaded only when a task calls for them, so your always-on instructions stay short. Ask your agent in plain language or name a skill by its ID.

Workflows

Debug systematically

Trace reproducible failures to evidence-backed root causes and verify fixes.

ap-debugDev flow

Run the dev flow

Take a goal through research, a reviewed and approved plan, reviewed task-by-task implementation, integration review, and verification in one worktree.

ap-dev-flowSecurity

Run a security audit

Inspect concrete attack surfaces, validate findings, and report actionable security risks.

ap-security-auditBackend

Develop APIs

Build thin, secure, documented API layers over reusable business and data logic.

ap-develop-apisFrontend

Design distinctive frontends

Create intentional, production-grade interfaces and maintain the project design system.

ap-frontend-designBackend

Write database queries

Write correct, secure, performant queries with sound transaction and indexing choices.

ap-write-database-queriesWorkflows

Start a development session

Orient a new session and isolate mutable local work in a dedicated Git worktree.

ap-start-dev-sessionPlanning

Create a product requirements document

Interview the user, shape the product, and write a clear evidence-aware PRD.

ap-create-prdManagement

Manage Agents Pack

Manage official and user-owned cross-agent components through the safe lifecycle CLI.

ap-manage-agents-pack

Browse all 42 skillsSee the core instruction set

Safety

It would rather refuse than guess.

Agents Pack edits files you already own, in repositories you already ship. Every mutation is planned, hashed, and reversible — and it prefers a clear refusal over a speculative merge.

  • Preview is honest

    --dry-run runs the real planner and writes nothing — no provider files, no installation state, no cache.

  • Unexpected edits stop the operation

    Agents Pack records a SHA-256 hash for every file it writes. If one changed underneath it, the command stops and tells you which file.

  • Shared files stay shared

    Codex instructions live in a marked block inside AGENTS.md. Everything outside that block is yours and is never touched.

  • Writes are transactional

    A failed operation restores the previous filesystem state, and rerunning an interrupted command recovers before it replans.

agents-pack status

Read-only. Reports the health of every managed output.

  • cleanMatches what Agents Pack last wrote
  • missingExpected managed content is gone
  • modifiedContent differs from the recorded output
  • malformedStructure or markers are invalid

Anything other than clean stops mutating commands until you resolve it. See Safety and drift.

Day-to-day use

You do not have to remember the commands.

Every installation includes the required ap-manage-agents-pack skill, which teaches your agent that the CLI is the source of truth and that generated files are never edited directly. Ask for the outcome you want.

  • Check my Agents Pack installation and explain whether everything is healthy. Do not change anything.

  • Install the ap-frontend-design skill. Preview the change first and wait for my approval before applying it.

  • Fork ap-debug into a user-owned skill named my-debug, then show me which canonical file I should customize.

  • Synchronize my user-owned Agents Pack components, then verify that every managed file is clean.

Open source

Nothing here is a black box.

Agents Pack writes instructions that shape how your coding agents behave. You should be able to read every word of that before you trust it, so all of it is public.

  • src/

    The CLI itself

    The TypeScript source for the planner, the writers, and the ownership checks — plus the test suite that proves them.

  • content/

    Every component

    The instruction set, all skills, and all subagents are plain Markdown. Read them, diff them, and see exactly what your agent is told.

  • install.sh

    The installer

    The same script the install command fetches. If you would rather not pipe anything to a shell, read it first and run it yourself.

  • registry/

    The release process

    How pack versions are published and resolved, so you can verify which content a given version actually contains.

Browse the source on GitHubFork a component and make it yours

Set it up once. Every agent benefits.

macOS and Linux, ARM64 and x64. One executable, checksum-verified, no sudo. Then run agents-pack init in your project.

curl -fsSL https://farfarawaylabs.github.io/agentspackai/install.sh | sh