Evaluating a Framework-Agnostic, Git-Native Agent Standard & CLI

Overview

An evaluation of an open-source, git-native standard (and its TypeScript CLI) for defining AI agents that port across Claude Code, OpenAI, CrewAI, Cursor, Gemini, Copilot, Lyzr and more via adapters. Studied as R&D into reusable, version-controlled agent definitions, not original studio authorship.

Why It Exists

Every AI framework has its own agent structure, so an agent built for one runtime rarely moves to another. We pulled this OSS project to assess its premise, “your repository becomes your agent”, and whether a single manifest plus Markdown identity files can replace per-framework wiring in our own work.

What We Built

This is an evaluation clone (~72 upstream commits), framed honestly. We examined the CLI’s command surface (init, validate, info, export, import, install, audit, skills, run, registry, lyzr) and its adapter layer, concrete adapters for claude-code, openai, codex, copilot, crewai, cursor, gemini, gitclaw, kiro, lyzr, nanobot, openclaw, and opencode. The standard centers on two required files (agent.yaml manifest, SOUL.md identity) with optional skills, tools, workflows, knowledge, memory, hooks, and compliance folders, including first-class FINRA/SEC/Fed compliance and segregation-of-duties modeling.

Technologies & Approach

TypeScript CLI built on Commander, Ajv (manifest schema validation), Inquirer, and js-yaml, published to npm. The interesting design is the adapter pattern that exports one definition to many runtimes, plus a system-prompt builder shared across adapters.

Outcome / Impact

Gave us a clear read on a maturing agent-definition standard and a well-factored multi-adapter CLI architecture, informing how we structure portable agents internally. Paired with the related architect assistant evaluation.

Capabilities Demonstrated

• Assessing agent-definition standards and their CLI tooling
• Understanding multi-framework adapter architectures (12+ runtimes)
• Manifest-schema validation and agent compliance modeling
• Honest evaluation of third-party OSS as capability research