DEV Community

Alain Prasquier
Alain Prasquier

Posted on

Announcing Org Agentic Toolkit (OAT): sane rule management for coding agents (Cursor, Claude Code, etc.)- Open Source

#python #ai #productivity #opensource

Open-source CLI to compile, validate, and ship consistent agent rules across every repo—deterministic, auditable, and IDE-ready.

Cursor, Claude Code, Windsurf, Continue, Copilot-ish variants… they all work — until they don’t. And when they don’t, it’s often not the model. It’s the rules:

  • “Do we run tests by default?”
  • “What’s our repo structure and conventions?”
  • “What’s allowed to be auto-refactored vs not?”
  • “How do we handle secrets, migrations, CI, code style, PR etiquette?”
  • “What does ‘done’ mean in this org?”

Right now, most teams manage this with:

  • copy/paste prompt blobs across repos
  • half-updated AGENTS.md files
  • tribal knowledge + “please read the README”
  • silent drift

So I’m releasing an open source tool that treats agent instructions like real configuration:

Org Agentic Toolkit (OAT)

OAT is a lightweight CLI that helps you manage and ship consistent rules for coding agents across an organization.

It’s built around a simple idea:

Org-level agent rules should be authored once, inherited explicitly, compiled deterministically, and validated.

What OAT gives you

  • One authoritative org “constitution” for coding agent behavior (security, style, testing, repo conventions, etc.)
  • Explicit inheritance per project (no mystery “where did this rule come from?”)
  • Deterministic compilation into a single output artifact (diffable + reviewable)
  • Validation so missing/invalid rule references fail fast (before your agent improvises)
  • Targets for IDE coding agents (Cursor today; more coming)

PyPI: https://pypi.org/project/org-agentic-toolkit/

GitHub: https://github.com/alain-sv/org-agentic-toolkit

The problem: prompt drift is a real engineering tax

If you have more than ~2 repos, prompt drift becomes inevitable:

  • repo A says “run tests”
  • repo B says “don’t run tests, too slow”
  • repo C says “use uv”
  • repo D uses poetry
  • half the team is on Cursor rules, half on Claude Code conventions, nobody knows what’s canonical

OAT makes the org policy canonical, and projects and users can add their own specifics without mutating the org baseline.

That last bit matters: personal preferences are great — but they shouldn’t silently weaken org safety rules.

Contributions welcome (seriously).

This is useful today, and it’ll be much better with community help.

If you’ve ever thought: “we need a source of truth for agent behavior” — you’re the target audience.

Repo: https://github.com/alain-sv/org-agentic-toolkit
Issues/PRs: very welcome.

⸻

Agents will write more of our code.

It’s time we managed their rules like we manage everything else: versioned, compiled, validated, reviewable.

Top comments (0)