A filesystem convention for storing project-level AI agent context inside a repository.
New to APC? Read the documentation.
Every AI coding tool today stores its agent state in its own folder. A project that uses two tools ends up with two folders, two agent definitions, two memory files — all drifting out of sync.
APC proposes a single, project-owned directory — .apc/ — committed alongside the code.
Any tool or runtime that implements the convention can read and write to it.
APC is a filesystem convention. It specifies:
- Project metadata — name, version, protocol version
- Agent definitions — role, model, skills, language, description
- Durable memory — per-agent
memory.mdthat survives across sessions - Session records — lightweight audit trail of what each agent did
- Skill registry — reusable prompt fragments agents can import
- Tool hints — optional MCP server declarations
APC does not define how runtimes execute agents, how models are called, or how UIs work. It defines only what a project stores on disk.
Model Context Protocol handles the wire protocol between agents and external tools. APC handles the project context layer. They operate at different levels and are designed to work together.