NevTan AI for VS Code

A local AI coding agent that works inside your editor. NevTan reads your workspace, writes and edits code, runs your builds and tests, reads the failures and keeps going — driven by a model running on hardware you control rather than a coding service you rent.

Ask for a change in plain English. NevTan finds the code, proposes the edit, shows you the diff, runs the tests, and fixes what it broke. You see every tool call it makes and approve anything that touches your files or your terminal.

What it does

Three modes, one conversation. Pick Agent, Ask or Plan from the composer. Agent is the full coding agent below. Ask answers questions about the project and cannot change anything. Plan investigates the project — including the state of the working tree — and writes an implementation plan, and cannot change anything either. Switching between them keeps the conversation, the transcript and the model you chose, so you can ask what something does, ask for a plan, and then switch to Agent to carry it out, all in one place. The limits are enforced where the tools run, not by hiding buttons.

Understands the code you are actually looking at. The agent is told which file is open, where the cursor is, what you have selected, which other editors are open and which branch you are on — and it is told plainly that this is a hint rather than an instruction. When you want something to be the subject of the question instead of context around it, attach it with +: a file, a selection, a symbol, or the problems the editor is already reporting.

Searches and reads the codebase. Full-text search across the project, directory listings, and file reads — so "why does the parser drop escaped quotes" is a question about your repository, not about programming in general.

Writes and edits code. Targeted edits to part of a file, or whole new files. Every change is shown as a diff against what was there before, and the turn ends with a list of the files it touched and how many lines moved in each.

Runs builds, tests and commands. With your permission, NevTan runs commands in your workspace, reads the output, and uses it. A failing test is not the end of a turn — it is the input to the next step.

Works through multi-step tasks. The agent loop plans, acts, reads the result and decides what to do next, reporting where it is as it goes ("step 6 of 24") so a long task is something you can wait through rather than wonder about. Stop it at any point; whatever it already changed stays changed, and the transcript says so.

Shows its work. Every tool call appears in the transcript as a sentence — "Searched Parser.java", "Ran mvn test" — folded away once the turn is done and one click from the full output. No log to go and read.

Permission, not trust

Reading and searching happen freely. Changing a file or running a command stops and asks, and the request names what it would do:

• See the diff before you answer. An edit approval opens VS Code's own diff view on the change that has not happened yet.
• Grant for the session when you get tired of approving the same kind of call — and see every standing permission listed in the composer, where you can take it back.
• The workspace is the boundary. NevTan works on the folders you have open. A file outside them is not read, not attached, and not edited.
• Changed files stay reviewable after the fact, including the ones a session grant let through without a prompt. The whole trade of granting permission is that the prompt goes away and the visibility does not.

Local models

NevTan runs open-weight coding models on your own machine. Turn on NevTan AI: Toggle Local Models on This Machine and the extension starts a bundled agent that downloads model weights and serves inference from them locally, with real progress as the weights come down and cancellation that actually reaches the download.

A conversation pinned to that device is answered by the model running there. Weights live on your disk, the inference happens on your hardware, and you can remove a model when you want the space back.

Local execution is opt-in and off by default. Installing this extension starts no native process on its own.

How your code is handled

Worth being exact about, because "local" is the reason to choose this.

NevTan coordinates a conversation through your NevTan account: the message you type and the context attached to it are sent to the NevTan API, which routes the turn to the model you selected. When that model is a local one, the weights and the inference stay on the machine running them — your code is not processed by a third-party AI provider, and nothing is trained on it.

nevtan.apiUrl points at whichever deployment you want, including one you run yourself.

Getting started

1. Install the extension and click NevTan AI in the status bar, bottom right. (Or run NevTan AI: Open NevTan AI from the command palette.) NevTan opens in the Secondary Side Bar, beside your code.
2. Sign in — approval happens in the browser you are already signed into. The extension never sees a password.
3. Open a folder. NevTan is a coding agent; it works on the files in a workspace.
4. Pick a model, or turn on local models and download one.
5. Ask for something.

Good first asks: "Explain how this project is laid out." · "Find out why the tests fail and fix the code, not the tests." · "Add a test for the edge case this function misses."

Requirements

• VS Code 1.106 or newer — NevTan lives in the Secondary Side Bar, and the contribution point for that arrived in 1.106
• A NevTan account
• For local model execution: Windows x64. The bundled inference agent currently ships for that platform only. Everything else — the agent loop, tools, approvals, diffs and conversations — works anywhere VS Code runs, with the model served by another device on your account or by the server.

Coding models are large. Expect a real download the first time.

Settings

apiUrl and webUrl must point at the same deployment: a sign-in started against one and approved on the other cannot complete.

Commands

Everything is under NevTan AI in the command palette. The ones worth knowing:

Building it yourself

See docs/ARCHITECTURE.md for how the extension, the bundled agent and the NevTan API fit together, and how to run it from source.