Guide Coding

No more vibe coding. Enter Guide Coding.

You write the code. AI guides you, one step at a time.

Guide Coding in VS Code: type a goal, get a plan of small steps, write each step yourself, and the AI reviews your attempt

Guide Coding is a VS Code extension that plans a coding task with the model of your choice and then walks you through it. It does not write the code for you. Each step shows up in the editor, right where the code belongs: what to do, why, and where.

The human writes, the AI guides. The model plans, explains, hints and reviews. You type.
Understanding is the goal. You finish a task knowing why the code is the way it is, not only holding code that works.
Generating code is a choice, not the default. Show code, Preview and Apply are there when you want them, one level at a time.

Your API key stays in your operating system's keychain and goes only to the provider you pick; nothing is sent to us. Want to try it without a key? Use a local model through Ollama, and nothing leaves your machine.

Preview. Guide Coding is an early release. It works, but expect rough edges. Please tell us what breaks.

Install

Search for Guide Coding in the Extensions view (publisher novian), or install it from the VS Code Marketplace (Open VSX: coming soon). Then open a folder you trust, and open Guide Coding in the activity bar.

Choose a provider

Guide Coding runs on your own model account. On first run, choose Choose provider (or run Guide: Choose model provider), pick one, and paste its API key. The key is checked for free (it only asks for the provider's model list), then stored in your operating system's keychain.

`guide.provider`	Your code is sent to	You need
`deepseek`	api.deepseek.com	a DeepSeek API key
`anthropic`	api.anthropic.com (Claude models)	an Anthropic Console API key
`openai`	api.openai.com	an OpenAI API key; you pick the model at setup
`openrouter`	openrouter.ai, and the vendor of the model you pick	an OpenRouter key; you pick the model at setup
`ollama`	nowhere: the model runs on your machine (localhost:11434)	no key; you pick one of your installed models
`custom`	the endpoint in `guide.baseUrl`	optional: a local server such as LM Studio takes no key

Anthropic with Claude Max or Team. Have Claude Max or Team? Your plan includes monthly API credits (Max 5x $100, Max 20x $200, Team up to $500 pooled). Claim them in claude.ai under Settings → Billing by linking a Claude Console organization, create an API key in that organization, and choose Anthropic. (Pro, Free and Enterprise plans don't include these credits.) Details.
Ollama and context length. Ollama's default context is 4k tokens on machines with less than 24 GiB of VRAM, and it silently cuts long prompts from the front. Guide Coding needs at least 32k. Raise it with the Ollama app's context slider, or set OLLAMA_CONTEXT_LENGTH (64k, 65536, is a good choice for coding tools). Pick a model that can call tools, for example ollama pull qwen3-coder.
Cost. For a model whose price is known (Anthropic's, DeepSeek's) the cost is shown in dollars, and the task pauses at guide.maxTaskCostUsd (default 5). For any other model, each turn's tokens are written to the log instead.

A custom endpoint

Any server that speaks the OpenAI Chat Completions API or Anthropic's Messages API works. Add three settings to your user settings (Preferences: Open User Settings (JSON)), then run Guide: Choose model provider and pick Custom endpoint:

guide.baseUrl: the API's base address. For the openai shape it ends just before /chat/completions, usually with /v1; for the anthropic shape it ends just before /v1/messages. It must be HTTPS, or plain HTTP only to localhost or 127.0.0.1: nothing is ever sent over plain HTTP to another machine.
guide.apiShape: openai or anthropic.
guide.model: the model's id exactly as the endpoint names it. Pick a model that can call tools: the guide works through tool calls.

A local model in LM Studio, with no key, so nothing leaves your machine (start LM Studio's server, and copy the model's API identifier from its Developer tab):

"guide.baseUrl": "http://localhost:1234/v1",
"guide.apiShape": "openai",
"guide.model": "qwen/qwen3-coder-30b"

Leave the key box empty when it asks.

A hosted OpenAI-compatible service, Groq for example:

"guide.baseUrl": "https://api.groq.com/openai/v1",
"guide.apiShape": "openai",
"guide.model": "<a model id from Groq's model list>"

Paste the service's key when asked; it is stored in your keychain and sent only to that host. The cost of a custom endpoint is not shown in dollars: each turn's tokens are written to the log instead.

Your key, safely

Where it lives. Keys are stored in your operating system's keychain (VS Code's secret storage), one per provider, and each is sent only to its own provider. Guide Coding has no server of its own and no telemetry.
Give it a key of its own, with a cap. Create a key just for Guide Coding and limit what it can spend: DeepSeek is prepaid, so your balance is the cap; OpenRouter keys take a credit limit; Anthropic Console workspaces and OpenAI projects take spend limits. Revoke the key at your provider whenever you like.
Change or remove it. Guide: Set API key replaces a key, and Guide: Clear API key deletes it from the keychain.

Your first task

Open a folder, then choose Start a task in the Guide Coding view and say what you want to build.
The model reads your code (read-only) and proposes a plan. For a bigger goal, the plan is split into milestones, each a slice that runs and shows something. The next milestone is planned only once you reach it, from the code as it is by then.
The current step appears in the editor. Type it yourself, then choose Check my work: the model reviews what you changed.
Stuck? Go one level further at a time:
- Hint: the shape of the code
- Ask: ask the model about the step, in the step's own balloon
- Show code, Preview (a diff of the file with the step applied) and Apply (one undoable edit)
At the end of a milestone, a checkpoint asks you to try what you built before moving on.

You own the plan: edit, skip, reorder, add and delete steps, and the model adapts. Two more tools for when a step does not fit:

Split step replaces the current step with smaller ones.
Finish later keeps your stub for now; a later step finishes it.

Learn or Control

Learn (the default) explains first: why the step is needed, how it works, where it goes.
Control is terse and shows the code at once, for when you know the ground and want the model out of your way.

Switch any time from the Guide Coding view's title bar, or with the guide.mode setting.

Privacy

Guide Coding collects nothing and sends nothing to its author; it has no telemetry. Your goal, the files the model reads, the diff of your changes when you Check, and your messages go only to the model provider you chose, under that provider's terms (with Ollama they stay on your machine). The model can read your workspace but cannot change it: only your Apply and Create file write, and a path guard keeps it out of anything outside the folder and secrets such as .env*, *.pem and .ssh. Your keys live in your keychain, and the task and its conversation in VS Code's workspace storage on your machine. Read the full PRIVACY.md; the terms of use are in the EULA.md.

Requirements

VS Code 1.101.2 or later (desktop).
A folder you trust, opened in VS Code. Untrusted and virtual workspaces are not supported, because the model reads your files.
An API key for one of the providers above, or a local model through Ollama.

Feedback

Found a bug, or have an idea? Open an issue at github.com/OnurGumus/guide-coding/issues. For a bug, please say your VS Code version, provider and model, and what you did. The Guide: Show log output helps, but remove any code and keys from it first.