Deep Code CLI

The deciding trait is focus. This is tuned for one model family rather than spread across a dozen, and that shows up as defaults you do not have to fight, prompts written for how that model actually behaves, and controls exposed where they matter instead of buried behind a generic abstraction.

You are the wrong buyer if you change providers with the weather, because everything good here is calibrated for a specific one. Pick it if your invoice already says DeepSeek. Pick a provider-agnostic terminal agent if it says something different every month.

It edits across files and it does not touch git. That combination means a session leaves your working tree changed with nothing marking where the changes began, and the recovery path is whatever you remembered to do beforehand. Every agent that writes to disk owes the user a boundary. This one leaves it as an exercise.

What it does right is stay small. It starts in whatever directory you are in, with no workspace concept of its own to configure, so nothing about your repository has to be registered before work begins.

1. Two of the headline controls are passthroughs. A deep-thinking toggle and an effort setting are parameters the provider already accepts, surfaced in a settings file, which is convenience rather than design. 2. The product claims optimisation for a specific model family and publishes nothing that would let a reader check what optimisation means here.

2. That is the gap worth naming: a tuning claim is an empirical claim, and an empirical claim without a measurement is a preference. The design will need revisiting whenever that model family changes shape.

2,225 stars is serious attention, and it was earned by attaching to a model family with its own momentum. That is a fine way to acquire users and a poor way to own them, because the thing driving the interest is not the thing this vendor controls.

Moat: none, and the closest thing to one is being first with defaults a provider could ship themselves in a release note. Likely acquirer: the model vendor, cheaply, or nobody. Position: free and useful today, and I would not model eighteen months of roadmap on somebody else's release cadence.

A global package install on sixty developer machines is a version-drift problem my endpoint team has solved before, so that part is routine. The part that is not routine is the destination: source code leaves for whichever inference endpoint is configured, and approving that destination is a procurement exercise, not a checkbox.

There is no SSO, no audit log, no retention policy and nothing that runs unattended, so it never becomes a measurable step. Zero licence cost, provider spend uncapped. Not yet, and the blocker is the data path.

The three settings that matter are plain and overridable from the environment: model, base URL, key. That means the tuning is a default rather than a cage, and I can aim the whole thing at any endpoint that speaks the common wire format without patching anything.

Skills load and MCP servers attach, so the capability surface is mine to extend rather than the vendor's to ration. MIT underneath all of it. The one thing I cannot do is serve the weights myself, which keeps a provider in the loop no matter how the variables are set.