The open Agent-Computer Interface runtime for real software

Uni-CLI

Name: Uni-CLI
Author: OLo

Find the operation. Cross the boundary. Keep the outcome inspectable.

Uni-CLI provides one searchable boundary between agents and 324 websites and tools, logged-in browsers, desktop apps, local commands, files, MCP servers, and system capabilities. It ranks executable operations by intent, runs the selected operation through its declared substrate under policy, returns a stable success/error envelope, and keeps failure diagnosable and repairable.

Intent discoveryDeclared substratesPolicy-awareStructured envelopesMCP + ACPBrowser + DesktopRepairable paths

Start in 30 seconds Browse operations

First commandv0.400.2 · Apollo · Duke

npm install -g @zenalexa/unicli
unicli search "find AI agent discussions on reddit"
unicli agents recommend codex
unicli mcp serve --transport streamable --port 19826

Positioning

One agent interface. Every useful software boundary underneath.

A model that can reason still needs an interface designed for bounded context: what can act, which substrate an operation declares, what the action can affect, what the call returned, and how to continue after failure. Uni-CLI does not replace the model or orchestrator. It organizes APIs, files, CLIs, browsers, desktops, protocols, and visual control into one discoverable, governed, observable, repairable runtime.

DiscoverBilingual BM25 search retrieves only the operations, arguments, auth posture, risk, and examples relevant to the task.

Select and governThe agent selects an operation with a declared strategy and substrate; capability scope, effect, risk, and approval remain inspectable before execution.

Act and observeThe adapter kernel invokes the selected operation; AgentEnvelope distinguishes success from error, and supporting operations add artifacts, recordings, or post-state evidence.

RepairErrors name the source path, failed boundary, retryability, suggestion, and alternatives, then verify a local repair or reroute.

Questions

Why it matters

What category is Uni-CLI?

An Agent-Computer Interface runtime: the executable boundary between an agent and real software. CLI is the native full process entry point; MCP projects adapter operations, while browser, desktop, and visual control are action substrates.

Why not just drive the browser directly?

The browser is one boundary. The catalog can hold API, file, CLI, page-semantic, CDP, desktop-accessibility, and visual operations. Today the agent selects the operation; the runtime does not arbitrate every alternative automatically.

What does an inspectable outcome mean?

Every rendered call distinguishes success from error in a stable envelope. Reads, file writes, browser mutations, and desktop actions add provenance, post-state, artifacts, or recordings only when that operation supports them. Dispatch cannot prove objective completion.

How does MCP fit?

MCP is one discovery and exposure substrate. Compact, deferred, and expanded profiles project adapter operations. Fixed core commands are currently canonical on native CLI; command-level parity is roadmap work.

Workflow

How intent becomes an inspectable result

Coverage

Current surface

These numbers come from the current generated repo artifacts: operations, adapters, pipeline steps, tests, and substrates are counted by the build.

sites and tools: 324
operations: 1817
pipeline steps: 105
tests: 9816

Surfaces

What it can control today

Web and communities

Public APIs, cookie sessions, RSS, search, downloads, publishing, and Chinese platforms.

Browser actions

CDP navigation, clicks, typing, intercepts, screenshots, snapshots, and before/after evidence.

Desktop and local

macOS AX, background input, Office, design tools, media tools, containers, and local subprocesses.

Agent protocols

MCP stdio / Streamable (legacy `sse` alias), ACP, agent matrix, skills export, and config generation.

Entrypoints

Start here

InstallInstall the CLI, run the first search, and learn output formats plus exit codes.
Operation catalogFind operations by site, substrate, auth mode, and examples.
Repair adaptersRead YAML, pipeline steps, the repair flow, and verification commands.
Connect agentsNative CLI, MCP, ACP, agent configs, and skills export.

Agent-readable index/llms.txt /llms-full.txt