Extension manager for coding agents and AI assistants
Extension manager for AI coding agents — install skills, subagents, MCP servers, rules, hooks, knowledge, and packs across Claude Code, Codex, Cursor, Copilot, and more.
Bundle extensions with packs to share workflows across projects, teams, and others.
Switch effortlessly between coding agents, bringing your workflow with you. Avoid lock-in.
[!WARNING]
AXM is in early alpha testing. Capabilities and APIs may change. To report
a problem or request a feature,
open an issue.
Multi-tool — a unified context-engineering toolkit for managing skills,
subagents, MCP servers, rules, hooks, knowledge, and packs
Multi-agent — share workflows across Claude Code, Codex, Cursor,
Gemini CLI, GitHub Copilot, OpenCode, and
nearly 40 others
Cross-platform — discover and distribute agent skills and extensions
across package ecosystems including JavaScript, Python, Java, .NET, Go,
Rust, Ruby, and more
Installation
Ask your agent to install AXM:
Follow these install instructions to set up AXM: https://axm.sh/install.md
Initialize AXM in your project. AXM detects your installed agents and creates
root axm.json; acquired packages and accepted external resolutions are kept
in agent_extensions/ and axm.lock, while .axm/ is runtime scratch.
AXM leaves Git attributes and formatter policy to the repository; exclude
agent_extensions/ from tools that rewrite files.
axm setup
See what's recommended for your project:
axm discover
Install an extension and AXM wires it into every agent in your workspace:
axm install @acme/skills/code-review
Extension types
Type
What it is
Governing standard
Skills
Package reusable agent skills with SKILL.md metadata and instructions.
Packs are the container type: a pack curates any of the above into one
bundle teams install and keep in sync, such as
@acme/packs/frontend-tools — the frontend guild's standard skills and
commands.
Every type has its own subcommand namespace (axm skills, axm mcps, axm subagents, axm rules, axm hooks, axm knowledge, axm packs) sharing a common shape: install, uninstall, list, update, new, publish, plus enable/disable where it applies.
Subcommands
Examples below use skills; the same shape applies to every type.
axm skills new my-skill # Scaffold a new skill
axm skills install @acme/skills/code-review
axm skills list
axm skills disable my-skill # Turn off without uninstalling
axm skills enable my-skill
axm skills import ./external-skill @acme/skills/my-skill
axm fork @acme/skills/code-review @me/skills/code-review-custom
axm skills publish # Publish authored skills
Installed and enabled skills are always materialized in .agents/skills/ for
the agentskills.io format, plus any declared agent-native skill directories.
Run axm --help for the top-level command list, or axm <command> --help
for options on any subcommand.
Workspace
Top-level commands work across every extension type and infer the target from
your input:
axm sync # Realize extensions declared by axm.json
axm install @acme/skills/code-review # Install a single extension
axm update # Pull latest versions
axm list # Inventory extensions across all types
axm list --outdated # Show extensions with available updates
axm list --deprecated # Show deprecated installed extensions
axm uninstall @acme/skills/code-review
axm adopt @acme/skills/retained-package # Make a canonical package authoritative
axm demote @acme/skills/review ./upstream # Explicitly return to external source management
axm lint # Report intrinsic workspace facts
axm sync --preview # Preview reconciliation without writing
axm sync # Reconcile desired, accepted, and observed state
axm upgrade # Update axm itself
Publishing
Extensions publish to the registry in four steps.
axm skills new my-skill # 1. Scaffold
# 2. Author content in the scaffolded directory
axm lint # 3. Check the publish gate locally
axm publish # 4. Publish new authored versions; skip published ones
Authorship is derived from the exact intrinsic workspace settings source,
the project owner, the settings map key, the extension type, and the matching
manifest in that type's authored root. Bare, filtered, and explicit selections
publish only workspace-authored packages. Adopt a retained canonical package
when this workspace should own it, or fork an installed package to publish it
under a new identity.
AXM builds a deterministic archive from each selected authored package whose
version the Registry does not have yet. A version the Registry already has,
yanked included, is skipped as already published without being rebuilt or
validated, so bump the version (axm version <fqn> patch) to ship local edits.
Installed external packages represent the source identity recorded
in axm.lock. axm lint checks their current contents against that accepted
identity. Publication reads workspace-authored packages.
axm lint checks the same rules the registry enforces — see
Lint for details.
Lint
axm lint evaluates workspace and per-extension invariants against a shared
rule catalog — the same one that drives the registry publish gate.
axm lint # Report findings against the current project
axm lint --fix # Apply every autofixable finding non-interactively
axm lint --scope user # Lint the user-scope workspace
axm lint --strict # Exit non-zero on warnings as well as errors
axm lint --json # Machine-readable findings envelope
Project scope is the default. Local lint.rules overrides in
axm.json affect axm lint only — the registry publish gate
remains authoritative.
Authentication
Sign in to a registry to publish, install private extensions, or attribute
usage to your account:
axm login # Sign in to the default registry
axm login --device-code # Sign in from SSH or a headless machine
axm whoami # Show the current identity
axm logout
axm token --output token # Write only the current token to stdout
axm login starts a local loopback PKCE flow, prints a manual authorization
URL, and then tries to open your browser. SSH without a display, CI,
Codespaces, and Linux with none of DISPLAY, WAYLAND_DISPLAY, or BROWSER
set (WSL excepted) automatically sign in with a code and say why; pass
--device-code to select that flow explicitly. Signing in from your terminal
asks you to open a link that already carries the code and check that the page
shows the same code; it also offers the plain page for typing the code by hand
and counts down until the code expires. It never touches your clipboard on its
own: press c while it waits to copy the link — over SSH, through the terminal
(OSC 52) to the clipboard of the machine you are typing on. Retrying the
command, or running axm login after you stop waiting, picks up the same
unexpired request; use --restart only when you intentionally want to replace
it. axm login --device-code --wait-for-human 300 starts or reuses the
request and waits up to 300 seconds in the same command; a bounded wait always
uses device-code sign-in.
axm token create --output token writes only a new token to stdout for a
pipe; axm help environment shows a guarded recipe for storing it.
A GitHub Actions job granted permissions: id-token: write needs no stored
secret: register its repository and workflow as a trusted publisher in
AgentXM settings, and AXM
exchanges the job's ID token for a short-lived workload token scoped to that
trusted publisher. AXM_TOKEN and AXM_TOKEN_FILE still take precedence, and
AXM_TRUSTED_PUBLISHING=0 turns the exchange off. See
Publishing from GitHub Actions
for a workflow, and axm help environment for the full credential order.
Supported agents
AXM supports nearly 40 agents — Claude Code, Codex, Cursor,
Gemini CLI, GitHub Copilot, OpenCode, Windsurf, Cline, Continue, Roo, Goose,
and more. If your agent isn't there,
open an issue and we'll add it.
Development
This is an Nx monorepo.
pnpm install # Install dependencies
pnpm build # Build all packages
pnpm test # Run tests
pnpm typecheck # Type check
pnpm lint # Lint
Use pnpm build:affected, pnpm test:affected, or pnpm lint:affected to
only operate on packages changed since main.
Extension manager for AI coding agents — install skills, subagents, MCP servers, rules, hooks, knowledge, and packs across Claude Code, Codex, Cursor, Copilot, and more.
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
Axm.sh ต้องใช้ API key ไหม?
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Axm.sh ใช้ทรานสปอร์ตแบบใด?
Axm.sh ทำงานแบบโลคัลผ่าน stdio
Axm.sh ถูกตรวจสอบล่าสุดเมื่อใด?
ตรวจสอบล่าสุดเมื่อ 2026-11-14 Extension manager for AI coding agents — install skills, subagents, MCP servers, rules, hooks, knowledge, and packs across Claude Code, Codex, Cursor, Copilot, and more.