Auggie CLI

You will get on with this quickly if you spend your day in a shell. The full-screen session streams output and shows each tool call as it fires, and that visibility is the trait that decides it in daily use, because an agent whose next move you can see is an agent you interrupt before it wastes ten minutes. Subagents, skills and custom commands are there when you want to shape it further.

Pick it when the editor extensions are already part of your team's habits. Pick Claude Code if you are starting fresh and want the larger ecosystem around your terminal agent.

The dependency is total. The row records that you cannot supply your own model and you cannot supply your own key, and the licence is proprietary, so there is no configuration, no fork and no fallback provider when something upstream changes. Claude, Gemini, GPT-5, GLM and Kimi are listed as backbones, but the routing between them is the vendor's decision and not yours. A tool with one supplier is a single point of failure wearing five model names.

What it does right: this ships as its own installable product rather than a mode bolted onto the editor extensions, so the terminal is a first-class surface.

Two observations. 1. Retrieval is the asserted differentiator, yet the documentation describes what the engine is for and never how it selects, ranks or bounds what it reads, so a reader cannot reproduce or falsify the claim. 2. No evaluation appears anywhere on the row, so there is no methodology to audit and equally no evidence to weigh.

Hooks, custom commands and typed SDKs indicate the loop is meant to be composed with, which is a design decision rather than an accident. The honest summary is that the surrounding machinery is specified in detail and the interesting part is not.

The interesting number is not the sticker. Usage bills as model inference plus compute with a forty per cent service fee on the inference, drawn against credit included in the plan, which means the company earns on volume it does not itself produce. That is real pricing power for as long as switching feels expensive, and it collapses the moment a customer notices they could buy those tokens directly.

The moat is the retrieval layer and the habit, not the models. Likely acquirer is a platform vendor that wants an enterprise context engine and a customer list. Position: fine as a supplier, poor as a dependency; keep an exit priced.

The demo is confident. Procurement is where it stops: the published plans each cover up to fifty seats, and I have sixty engineers, so I am in a custom Enterprise conversation before I have a figure to take to finance. The row records nothing about single sign-on, SCIM or retention, which means my security questionnaire goes out unanswered. Windows engineers get this through WSL only, which is a support ticket I can already picture.

In its favour, a print flag gives me a scriptable run, so pipelines are genuinely possible. Approved with conditions: a written SSO and audit commitment in the contract, or no signature.

I install it with npm and after that I own nothing. The licence is proprietary, so there is no source to read and no fork to keep alive if the company loses interest. My key is not accepted and neither is my hardware, which rules out the offline setup I actually work in. Grudgingly: it speaks MCP in both directions, and auggie --mcp turns it into a server other applications can call, which is more openness than closed tools usually offer.

That flag is the one place I can bend it. Everything else is a menu.