An open-source, community-driven dependency manager for the agent supply chain.
One manifest, one lockfile and one org policy for the skills, plugins and MCP servers your agents load, with the harnesses and marketplaces you already use.
GitHub Copilot | Claude Code | Grok Build | Cursor | OpenCode | Codex | Gemini CLI | Windsurf | Kiro
Documentation | Quickstart | CLI Reference | Roadmap
Portable by manifest. Verified by lockfile. Governed by policy. One file declares what your agents load; one lockfile reproduces it on every machine and in CI; one policy decides what your organization allows.
What a coding agent loads decides what it knows and what it can do: skills, instructions, prompts, agents, hooks, plugins and MCP servers. Today teams copy those files between repositories and hand-edit MCP config for every tool. The copies drift, nobody can say which version a repository follows, and nothing reproduces the setup on the next laptop or in CI.
APM treats them as dependencies. Declare them once in apm.yml. apm install writes each one in every harness's own format and records the exact commit, plus a SHA-256 hash of every file it wrote, in apm.lock.yaml. Think package.json and package-lock.json, or Cargo.toml and Cargo.lock, for your agents, with transitive dependency resolution included.
# apm.yml -- ships with your project
name: your-project
version: 1.0.0
targets: [copilot, claude] # the harnesses this project supports
dependencies:
apm:
# A skill, picked from a public collection
- git: vercel-labs/agent-skills
skills:
- deploy-to-vercel
# A plugin, from Anthropic's marketplace
- git: https://github.com/anthropics/claude-plugins-official
path: plugins/commit-commands
ref: main
mcp:
# An MCP server, configured for every harness above
- name: io.github.github/github-mcp-server
transport: http # MCP transport name, not URL scheme -- connects over HTTPSgit clone <org/repo> && cd <repo>
apm install # the same skills, plugins and MCP servers, from the lockfileComing from npx skills add? Drop-in:
apm install vercel-labs/agent-skills # whole bundle, like npx skills add
apm install vercel-labs/agent-skills --skill deploy-to-vercel # one skill, persisted to apm.ymlSame install gesture. You also get a manifest, a lockfile and a fresh clone that matches yours. Coming from Claude Code marketplaces, gh skill or hand-copied folders? See Coming from another tool.
Plugins from the marketplaces you already use:
apm marketplace add anthropics/claude-code
apm install commit-commands@claude-code-pluginsOne plugin, written in each harness's own format (commands for Claude Code, prompts for GitHub Copilot) and pinned in the lockfile.
MCP servers without hand-edited JSON:
apm install --mcp io.github.github/github-mcp-server --transport httpEvery harness keeps MCP servers in its own file: .mcp.json for Claude Code and GitHub Copilot, .cursor/mcp.json for Cursor, .codex/config.toml for Codex. Declare a server once and APM writes each file, on every machine. Add a harness later and the same servers follow. Tokens stay out of the repository as ${VAR} references, and a server that arrives through someone else's package with its own command is held back until you declare or approve it.
One apm.yml declares every primitive your agents need -- instructions, skills, prompts, agents, hooks, plugins, MCP servers -- and apm install reproduces the same setup in every harness the project targets, on every machine. No account and no registry: packages come from git.
- One manifest, every harness you target -- declared once, written for Copilot, Claude Code, Grok Build, Cursor, OpenCode, Codex, Gemini CLI, Windsurf and Kiro
- Skills, one at a time or whole collections -- from GitHub, GitHub Enterprise, Azure DevOps, GitLab or a plain git URL
- Plugins and marketplaces -- install Claude Code, GitHub Copilot and Codex plugins by name, locked like any other dependency
- MCP servers -- from the GitHub MCP Registry, your own registry, a local command or a remote URL
- Transitive dependencies -- packages can depend on packages; APM resolves the full tree
- Publish packages -- tag a release and consumers pin
your-org/your-package#v1.2.0; fixes reach them as Dependabot or Renovate pull requests - Build plugins -- keep your plugin's own dependencies declared and locked;
apm packwrites theplugin.jsoneach tool expects, including Codex's.codex-plugin/plugin.json - CI/CD ready -- the GitHub Action installs from the lockfile on every run
Agent context is executable in effect -- a prompt is a program for an LLM, and hooks and MCP servers start processes on your machine. APM records exactly what you installed and checks it on every clone and every pull request. A lockfile records what arrived, not that it can be trusted: review a new dependency, or a version bump, the way you review code.
- Lockfile integrity --
apm.lock.yamlpins each source to a commit and records a SHA-256 hash per package and per deployed file; a download that does not match is refused - Reproducible installs --
apm install --frozeninstalls exactly what the lockfile records and fails whenapm.ymldisagrees - Drift detection --
apm auditrebuilds what the lockfile describes and diffs it against your working tree;apm audit --cimakes it a pull-request check - Content scanning -- every install scans for hidden Unicode that can carry instructions a reviewer cannot see, and blocks critical findings; it looks for hidden characters, not intent
- Approvals for hooks and MCP servers --
apm approverecords, inapm.yml, which packages may deploy executables, so the grant is reviewed like any other change - Inventory export --
apm lock export --format cyclonedx|spdxlists what reached disk, straight from the lockfile -- an inventory for procurement, not an attestation
Teams that run coding agents at scale end up building the same thing: a central team publishes approved skills, plugins and MCP servers through its own marketplace, and many repositories install from it. apm-policy.yml lets that team say which sources, MCP servers and executables the organization allows, inherited from an enterprise baseline that each organization can tighten but never loosen, and a required apm audit --ci check enforces it on every pull request. Policy starts in warn mode; you switch it to block when you are ready. That is the dependency layer for an agentic SDLC.
APM governs what gets installed; your harness governs what runs. Policy is not a sandbox.
- Build your organization's marketplace -- publish approved capabilities as versioned packages that every repository installs by name and updates on purpose
- Set an org policy -- allow or deny sources, MCP servers and executables, in warn or block mode
- Gate pull requests -- wire
apm audit --ciinto a required check, so a denied source fails the build - Security model -- every control, its default, how to tighten it, and what stays your responsibility
- Roll out in stages -- from lockfiles to a blocking policy across hundreds of repositories, each stage reversible
brew install apmNo custom tap is required. Homebrew manages installation and updates:
use brew upgrade apm, not apm self-update.
curl -sSL https://aka.ms/apm-unix | shirm https://aka.ms/apm-windows | iexNative release binaries are published for macOS, Linux, and Windows x86_64.
For standalone installations, apm self-update reuses the matching platform installer.
You do not need Homebrew to use APM.
Other install methods: pip, WinGet, and Scoop
# pip (Python 3.11+)
pip install apm-cli# WinGet
winget install --id Microsoft.APM --exact --source winget# Scoop
scoop install apm
# pip
pip install apm-cliAlready using the Microsoft Homebrew tap? See the migration guide.
Then follow the Quickstart: add a skill, a plugin and an MCP server, commit apm.yml and apm.lock.yaml, and check a teammate's fresh clone against the lockfile, in five minutes.
Teach your coding agent to use APM with the official guide (an expert agent plus the apm-usage skill):
apm install microsoft/apm/packages/apm-guide#v0.37.0Created by @danielmeppiel. Maintained by @danielmeppiel and @sergio-sisternes-epam.
- Roadmap
- Discussions
- Contributing
- Adopters -- organizations using APM, listed with their consent; add yours
- AI Native Development guide -- a practical learning path for AI-native development
In the ecosystem: Thoughtworks Technology Radar Vol. 35 places APM in Trial, describing it as a tool that treats agent context as a supply chain. Dependabot and Renovate update APM manifests; JFrog Artifactory offers Agent Packages repositories. The formats are an open specification, OpenAPM v0.1, with conformance tests.
Built on open standards: AGENTS.md | Agent Skills | MCP
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.