Why it matters
As AI agents become more sophisticated, their ability to interact with real-world software is crucial. Uni-CLI addresses this by providing a robust and reusable execution layer that goes beyond simple API wrappers. Its focus on policy, evidence, and self-repair capabilities could significantly enhance the reliability and safety of AI agent operations across diverse computing environments, making agents more effective in complex tasks.

Uni-CLI, developed by olo-dot-io, is an open-source project that serves as an operations substrate for AI agents. It enables agents to interact with a wide array of software, encompassing 311 sites/tools, logged-in browsers, desktop applications, and local tools. The project aims to provide a consistent execution layer that agents can reuse across different runtimes, particularly when dealing with software drift.

Key features of Uni-CLI include the ability to turn intents into ranked, inspectable command plans, offering 'unicli search' and 'unicli do' functionalities. For safety, it incorporates permission profiles with 'open', 'confirm', and 'locked' execution modes. Every operation generates an 'AgentEnvelope' containing data, context, retryability information, and evidence hooks. The 'unicli delivery' feature uses this evidence to diagnose failures, suggest next experiments, and manage repair bounds.

Uni-CLI supports local computer use, including macOS AX, UIA/AT SPI sidecars, subprocesses, and visual input. It also facilitates exposing capabilities to agents via 'unicli mcp serve', ACP, native CLI, and JSON streams, all sharing the same command catalog. The project emphasizes turning useful operations into typed, searchable commands, enforcing policies for dangerous actions, returning machine-readable receipts, and making failures repairable by pinpointing the exact adapter or pipeline step that failed. It is written in TypeScript and released under the Apache License 2.0.

Share:XHacker NewsLink
Article ID - cmpjwpbnz0