The package manager for AI agents
Package the building blocks of agent systems once, version them cleanly, and make them portable across runtimes, frameworks, and ecosystems.
brew tap agentpm-dev/tap
brew install agentpmagentpm install @ns/package@0.1.0agentpm run @ns/toolname --input '{"text":"hello world"}'If your goal is to get to a working agent or workflow quickly, start with a template. It gives you a scaffolded project, wired dependencies, and clear next steps before you ever need to think about lower-level package composition.
Start with a workflow template
Templates are the fastest path to a working agent or workflow. Instead of assembling packages by hand, bootstrap a project with code, docs, local manifests, and runnable AgentPM dependencies already wired together.
agentpm new @ns/workflow-template@0.1.0 my-projectMinimal AgentPM Harness starter that installs one published Agent package and runs it through the reference Harness TUI or headless CLI.
AgentPM Harness starter that hosts a local embedding provider for the published AgentPM docs Knowledge package.
AgentPM Harness starter showing Harness-managed MCP exports and scoped runtime MCP imports in one workspace.
AgentPM Harness lab for process-hosted model, embedding, KnowledgeRuntime, and MemoryRuntime providers.
Python SDK starter for a local triage worker that combines a published AgentPM agent with one extra direct tool.
Python SDK starter that runs AgentPM Harness through a hosted model, Hook, and approval controller.
Node SDK starter that runs AgentPM Harness through a hosted model, Hook, and approval controller.
Multi-agent workspace starter that combines one published agent root, a support loop, support knowledge, support memory, one published support instruction profile, and local generated agent manifests.
Templates get you to a working project fast. When you want more control, packages are the lower-level building blocks you can inspect, install, swap, and compose directly.
Portable capabilities
Start with the concrete actions your agent needs to take. Tools are the executable building blocks you compose into larger systems.
agentpm install @ns/toolname@0.1.0Read and update GitHub issues using a compact, agent-friendly action surface.
Apply deterministic transformation operations to JSON objects and arrays.
Query CSV data with structured filter, sort, select, and aggregate operations.
Summarize long text to a concise abstract.
Translate text into a target language using LLM.
Sentiment analysis using VADER.
Reusable procedures
Once the actions are in place, package the know-how around them. Skills capture the playbooks, checklists, and reasoning guides that turn tools into repeatable workflows.
agentpm install @ns/skillname@0.1.0Procedural guidance for reviewing issue queues, identifying priority, and drafting follow-through.
Workflow guidance for turning operational findings into concise Slack updates.
Operational handoff checklist for incident updates, shift changes, and follow-through tracking.
Multi-step workflow guidance for gathering sources and producing a concise research brief.
Prepared retrieval context
Package the context your systems need to read from repeatedly. Knowledge captures curated documents and vector-ready corpora as portable, inspectable dependencies.
agentpm install @ns/knowledgename@0.1.0Prepared vector-mode retrieval corpus built from the public AgentPM v0.1 docs set.
Small vector-mode maintainer guide for issue triage, labeling, and review-comment drafting workflows.
Multi-file support response handbook for direct context loading across drafting, escalation, and customer messaging workflows.
Structured memory models
Define what durable memory should look like before you bind it to any runtime store. Memory Blueprints package record shapes, retrieval modes, and lifecycle policy as portable contracts.
agentpm install @ns/memoryname@0.1.0Simple support-facing Memory Blueprint for one durable customer state document with light lifecycle policy.
Workflow-oriented Memory Blueprint for maintainer preferences, active work threads, and follow-up notes across repository review sessions.
Flagship multi-space Memory Blueprint for durable conversation continuity across recent interactions, summaries, and saved notes.
Authored behavior layers
Package durable role, objective, and communication guidance as a portable layer of authored behavior. Instruction Profiles make that guidance inspectable and reusable across systems while keeping it distinct from execution logic.
agentpm install @ns/profilename@0.1.0Portable maintainer profile for concise repository triage, review guidance, and explicit approval-minded follow-through.
Portable incident-operations profile for calm status reporting, explicit risk framing, and structured handoff communication.
Portable support-response profile for calm, direct customer communication and explicit next-step handling.
Declarative orchestration
Package iterative control flow as a reusable graph with phases, transitions, checkpoints, and failure policy. Loops make the orchestration contract inspectable before any runtime decides how to execute it.
agentpm install @ns/loopname@0.1.0Portable ReAct loop for reasoning over context, taking actions, reflecting on results, and iterating until completion or handoff.
Portable maintainer-review loop for inspecting issues, drafting follow-up comments, and deciding whether to resolve, revise, or hand off.
Portable incident-response loop for assess, execute, review, and revision-driven operational follow-through.
Portable support orchestration loop for triage, response drafting, and escalation handoff decisions.
Tools, skills, knowledge, memory, profiles, and loops are the parts. An Agent is the thing you actually run: one portable package that resolves its own dependencies and executes through the Harness.
Install it, then run it
agentpm harness starts a real run: resolved capabilities, live model calls, approvals where the Agent asks for them, and a traced result you can inspect afterward.
agentpm install @zack/devwork-copilot@0.1.6
agentpm harness @zack/devwork-copilot@0.1.6Interactive GitHub maintainer copilot with explicit approval gates for write actions, packaged maintainer-review orchestration, workflow continuity, and a durable maintainer behavior profile.
Operations assistant for incident assessment, execution, and review workflows, with packaged continuity memory, incident-response orchestration, and durable incident-operator communication style.
Research assistant for crawling public sources, extracting structured content, and producing concise summaries.