Playbooks · Coding tasks
One manifest for every harness: APM, explained
Written November 9, 2026
APM promises one apm.yml manifest for every agent harness — skills, MCP servers, hooks, subagents. What it covers, what it doesn't, and the May date.
The problem: the same setup, maintained five times
A developer running several agent tools maintains the same skills, MCP servers, hooks, and subagents in every one of them — once for Copilot, once for Claude Code, once for Cursor, and so on. Each setup is hand-curated, undocumented, and drifts. Microsoft's Agent Package Manager (APM) is a dependency manager for that configuration: one apm.yml manifest in the project, and apm install deploys it to every tool.
What one manifest covers
An apm.yml declares instructions, skills, prompts, agents, hooks, plugins, and MCP servers, and APM writes them into the directories each tool already reads — .github/ for Copilot, .claude/ for Claude Code, .cursor/ for Cursor, and the equivalents for OpenCode, Codex, Gemini, Windsurf, Kiro, and Grok Build. Dependencies resolve from any git host, with versions pinned by tag and checked by content hash. A generated apm.lock.yaml pins the resolved tree the way package-lock.json does for npm, so a fresh clone reproduces the same setup.
Three mechanisms guard the supply chain: apm install scans package content for hidden Unicode that could hijack agent behaviour, the lockfile records content hashes for provenance, and apm audit rebuilds the expected configuration and diffs it against the working tree to catch hand-edits. An apm-policy.yml lets a team say which sources, scopes, and primitives are allowed, with tighten-only inheritance from enterprise to org to repo — a repo cannot relax a constraint set above it. The policy governs what gets installed; the agent tool governs what runs.
The date correction: May, not November
Stories in early November described APM as newly released. That is wrong. APM reached v0.12.2 on May 5, 2026, and the project dates to the previous autumn. What actually happened in November is narrower: Dependabot added version updates for dependencies declared in apm.yml, and the project drew fresh attention. Carry May as the date.
Verdict: use APM if your team runs two or more agent tools and wants identical instructions across them. Skip it if you are solo on one tool — the manifest and lockfile are overhead when one AGENTS.md already does the job.
What it does not do
- No remote MCP for Codex CLI. Codex CLI does not yet support remote MCP servers over streamable HTTP, so APM skips those declarations for that target with a notice. Local MCP servers still work.
- No runtime protection. APM defends the build-time supply chain — what gets installed and read. It does not sandbox MCP servers at runtime, does not perform malware analysis on dependency code, and does not sign packages.
- Not every primitive lands on every target. The manifest schema is a working draft, and support varies by tool — check the targets matrix before assuming a hook or MCP declaration reaches all of yours.
- Policy needs enforcement to bite. Without CI wiring such as
apm audit --ciin branch protection, anapm-policy.ymlis advisory.
Limits
- APM is young software, and its manifest schema is a working draft that can change. Pin versions and keep the lockfile committed.
- Dependabot coverage is version updates for
apm.ymldependencies, not security updates. - This page describes the tool from its repo and docs. No manifest was installed and run for it here.
Related playbooks
- Check every package name your agent invents
- Vera: unlimited AI code review, free at every tier
- Dependabot now updates agent package managers
- Check whether your Claude Code hooks are adding tokens to every prompt or cutting them
Sources
- Agent Package Manager — microsoft/apm on GitHub
- Manifest Schema — Agent Package Manager docs
- Microsoft APM: The Package Manager for AI Agents — Codex Knowledge Base
Back to: Coding playbooks hub · Also: Dependabot now updates agent package managers · Agent Skills has a neutral home — and still no security story · Pick a memory server for your coding agent