What is Lybra

Lybra is a local-first, file-authoritative control plane for AI agents — a harness that puts governance, verification, and observability first.

Most agent platforms optimize for autonomy. Lybra optimizes for accountability.

• Files are the single source of truth — state lives in durable files, not in compressed conversation memory. State outlives the model.
• The owner holds every decision gate — architecture, risk, and scope decisions are non-delegable.
• Independent audit is enforced — an executor can never audit its own work.
• Agents come to Lybra (MCP-native) — no autonomous runtime, no heartbeat polling.

The model isn't the bottleneck. The harness is.

Why it matters

Enterprises need AI they can hold accountable — not a polished demo. If you can't assign responsibility, reproduce a result, or audit a decision, you can't put an agent into real business. Lybra moves AI agents from demos to accountable, repeatable work.

How it works

Three peer surfaces share one permission-and-audit backend:

| Surface | Role | |---------|------| | CLI | command-line operation | | Board | local dashboard for review | | MCP | any agent connects in |

Every write to the workspace is forced through one path — no shortcuts:

dry-run  →  confirm  →  Owner Decision Gate  →  independent audit  →  written to files

Where Lybra sits — ETCLOVG

Against the seven-layer harness model (Execution · Tooling · Context · Lifecycle · Observability · Verification · Governance), Lybra is deliberately heavy on Governance, Verification, and Observability, and deliberately does not do autonomous Execution or Lifecycle orchestration. That's a stance, not a gap.

See docs/positioning/etclovg_self_assessment.md.

Install

npm install -g lybra
lybra --help

Lybra needs Node.js 18+ and Python 3 available on PATH.

For source checkout contributors:

git clone <repo-url>
cd lybra
python3 -m unittest discover -s tools/aipos_cli/tests
python3 -m unittest discover -s web/board/tests

Where the command needs a workspace root, use AIPOS_WORKSPACE_ROOT or --repo-root as applicable.

Quick start

lybra --help

The closed loop

For complex work, every task runs the same accountable loop:

Plan  →  Execute  →  Independent Audit  →  Fix / Finalize

The executor and the auditor must be different parties; nothing is finalized without an audit pass. Simple tasks close in a single step — but still leave an auditable trail.

Contributing

Changes that affect workflow gates, persistence boundaries, default ports, release surfaces, or audit rules should stay within the Owner decision flow and be validated before finalize. Keep product changes file-authoritative and narrowly scoped.

About

Built by KIWIAI.

Lybra helps teams turn AI from a personal trial tool into a manageable, reusable, traceable, and continuously improvable enterprise-grade system.

License

Apache-2.0. See LICENSE.