Open Agent Manifest v0.1
A portable manifest standard for AI agent packages.
Open Agent Manifest v0.1 is a small YAML contract for describing agent packages without forcing a runtime. It captures origin, protocols, permissions, safety posture, distribution URLs, and install instructions.
Manifest-first
The registry stores and displays manifests/packages. It does not host agent execution, runtime sessions, prompts with credentials, or tool secrets.
Framework-neutral
Listings can originate from LangGraph, CrewAI, Mastra, OpenAI Agents SDK, prompt-only workflows, MCP servers, or A2A endpoints.
Reviewable by design
Permissions, risk level, human approval requirements, install method, trust labels, and signature status are visible before import.
External execution
Execution happens manually in AI apps, locally, through MCP tools, or through A2A-compatible endpoints chosen by the user.
Required fields
The standard starts intentionally small. It should be readable by humans, easy to validate, and usable by many agent ecosystems without forcing one runtime.
Example YAML
open_agent_manifest: 0.1 name: Weekly KPI Brief slug: weekly-kpi-brief version: 0.1.0 agent_type: local_runtime framework_origin: Mastra model_requirements: - GPT-4.1 - Claude 3.5 Sonnet supported_protocols: - open_agent_manifest_v0.1 - local_package distribution: manifest_url: https://manifests.openagentsource.com/weekly-kpi-brief.yaml package_url: https://packages.openagentsource.com/weekly-kpi-brief.zip permissions: - analytics.read - sheets.read - docs.write_drafts safety: risk_level: low human_approval_required: true install: method: local_package instructions: Configure metric sources locally and approve the draft before publishing. execution: provided_by: external note: OpenAgentSource validates and distributes manifests only.