agentdef 0.2.0

Every AI platform invents its own way to define agents - CLAUDE.md, AGENTS.md, cursor-rules.md, copilot-instructions.md, declarative manifests... The concepts are always the same; only the files differ. AgentDef is an open specification that captures what an agent IS - identity, instructions, skills, workflows, tools - in one canonical, human-readable directory. Adapters generate the file each platform wants; importers convert your existing files back into the canonical form (505 real community agents imported with zero failures).

Quick Start

An AgentDef agent is just a directory. The spec defines canonical components - agent.md (identity), manifest.yaml (composition), instructions/, and optionally skills/, workflows/, tools/, knowledge/, memory/, runtime/, evals/ - everything is plain markdown and YAML, made to live in git.

The smallest valid agent:

my-agent/
├── agent.md          # Who is this agent?
├── manifest.yaml     # What does it use?
└── instructions/
    └── core.md       # How does it behave?

Copy the starter template and fill in your content - or let agentdef init do it for you:

CLI

Install AgentDef:

pip install agentdef

8 adapters (Claude, OpenAI AGENTS.md, Cursor, GitHub Copilot, LangGraph, M365 Copilot, OpenAI Assistants, CrewAI). 9 importers (those plus Copilot Studio, Letta .af, and any markdown prompt file).

agentdef init my-agent --yes                 # scaffold a valid agent
agentdef validate ./my-agent/
agentdef adapt claude ./my-agent/ --output CLAUDE.md
agentdef import copilotstudio agent-doc.md --output ./my-agent/
agentdef sync ./my-agent/                    # regenerate all configured framework files
agentdef list                                # every adapter/importer framework

Every import writes an IMPORT_REPORT.md: what mapped, what was inferred, what was dropped - nothing is dropped silently.

Take the tour

Explore this repo's own architecture as an interactive knowledge graph, with a guided 10-step walkthrough of the codebase:

▶ Open the live dashboard - generated from the code itself with the understand-anything plugin, so it can't silently go stale.

Documentation

Full docs: https://agentdef.github.io/agentdef/

• Getting started - first agent, validate, adapt, sync
• Migrate in 5 minutes - from CLAUDE.md, Copilot, Cursor, AGENTS.md, or any prompt file
• Examples - including a real end-to-end demo (Claude subagent → AgentDef → Copilot)
• Importer scorecard - Comparisons - FAQ
• Specification - spec 0.5, with a public conformance corpus

Status

Spec 0.5.0 - tooling 0.2.0 - see STATUS and the ROADMAP. The format is stable enough for experimentation and feedback; breaking changes are possible before 1.0.

Contributing

See CONTRIBUTING.md. Want a framework we don't cover? Open a framework request.

License

Apache 2.0 - see LICENSE.