Introduction

twinny is an AI coding assistant for Visual Studio Code that runs against a model server you choose. It was built for people who want the useful parts of a Copilot-style assistant, code completion, chat about their code, edits and reviews, without sending their code to a service they do not control and without a subscription.

Principles

Local first. twinny is designed around servers that run on your own hardware: Ollama, LM Studio, llama.cpp and anything that speaks the OpenAI API. Hosted APIs are supported too, but nothing in twinny requires one.

Nothing leaves your machine unless you point it somewhere. There is no account, no telemetry and no relay. Requests go to the server address you configured and nowhere else. The one networked feature that is not a plain HTTP request, Devices, connects your own machines to each other over an encrypted peer-to-peer link.

Small, explicit features. twinny does not run an autonomous agent over your repository. Every feature is a single thing you ask for and can see the result of: a suggestion you accept with Tab, an edit you review as a diff, a command you confirm before it runs. This is deliberate: it keeps the tool predictable, and it works well with the small models that fit on a laptop.

Works with weak models. Prompts are short, context is chosen carefully, and output is cut at sensible boundaries, so a 1.5B or 7B model gives useful results.

What you get

Feature	In one line	Learn more
Code completion	Ghost-text suggestions as you type, from a fill-in-the-middle model	Code completion
Inline edit	Ctrl+I, describe a change, review it as a diff	Inline edit
Chat	Ask about your code in the sidebar, with @ mentions for context	Chat
Context and mentions	Files, symbols, problems, git, terminal output and workspace search	Context
Workspace index	Embeddings plus keyword search plus a local reranker	Workspace index
Code review	Review the working tree, a branch, or a GitHub pull request	Code review
Commit messages	One click in the Source Control view	Commit messages
Terminal	Write a command from a description, fix the last error	Terminal
Devices	Use another of your computers’ GPU over P2P	Devices
Prompt templates	Every prompt is a Handlebars file you can edit	Templates

How it fits together

 VS Code ──► twinny extension ──► provider ──► model server
                 │                  (chat, autocomplete, embeddings)
                 ├─ sidebar webview (chat, providers, review, index)
                 ├─ inline completion provider (ghost text)
                 ├─ inline edit (diff regions, CodeLens, code actions)
                 ├─ workspace index (LanceDB + local reranker)
                 └─ P2P runtime (gateway to paired devices)

A provider is one model on one server for one job. You can have several and switch between them. The three jobs, chat, autocomplete and embeddings, want different kinds of model; the Providers overview explains which.

Requirements

• Visual Studio Code 1.93 or newer, or a compatible build such as VSCodium.
• A model server: something local (see Choosing a model server), a paired device, or a hosted API key.
• For the terminal features, VS Code’s shell integration, which is on by default for bash, zsh, fish and PowerShell.

Licence

twinny is open source under the MIT licence. The code is at github.com/twinnydotdev/twinny.