Kiln for VS Code

Kiln for VS Code brings your Kiln projects into the editor and adds a coding assistant that works on any folder on your machine.

The extension operates in two distinct modes, determined by the folder you have open. Which mode applies is stated in the Kiln view at all times.

Requirements

• Visual Studio Code 1.85 or later.
• A Kiln account, for working on Kiln projects.
• For the direct assistant, an API key from any supported provider. Keys are stored in the editor's secret storage and are never transmitted to Kiln.

Getting started

The quickest route is Open in VS Code, available from the ⋯ menu on any project in your Kiln dashboard. It installs nothing further, prompts for a destination folder, and downloads the project.

To connect manually:

1. Run Kiln: Sign In from the Command Palette (Ctrl+Shift+P or ⌘+⇧+P). An access token is recommended over an email and password: it is named, can be given an expiry, and can be revoked independently, so a lost machine costs you the machine rather than the account.
2. Run Kiln: Open Project… and choose a project and a destination.
3. Edit the files as you would any other project, then run Kiln: Push Project, or use the status bar.

A downloaded project is an ordinary static site — pages, styles/, scripts/, a README.md and a package.json. npm run dev in that folder serves it.

Working on a Kiln project

Pushes create versions, never overwrites. Every push is stored as a new version with a label, reachable from Kiln: Show Version History… here and from the project's history in the browser.

Changes can be read before they are sent. The push dialog offers a diff against what is stored at that moment, rather than against the copy you originally downloaded — the two differ precisely when a collaborator has been working, which is when it matters.

The assistant proposes; you accept. Kiln: Open Assistant reads the project first, including which files are shared between pages, then proposes an edit. Nothing is stored until you accept it, at which point it lands as a single labelled version through the same path as your own pushes.

Conversations are shared with the browser. The thread lives on the server, so it is the same conversation in the editor and in the web app. Closing the panel mid-response does not interrupt it.

Guardrails are enforced by the server. The rules governing what may be stored do not live in this extension, so a defect here cannot damage a project.

Working on any other folder

In a folder that is not a Kiln project, the assistant becomes a direct coding agent. The differences are material and worth stating plainly.

• Requests go straight to the model provider. The Kiln server is not in the request path, and nothing you send reaches it.
• There is no staging step. Files are created and modified immediately, and shell commands are executed without individual confirmation. Use this mode on folders where that is appropriate.
• There are no imposed limits — no round cap, token ceiling, or command timeout. Your provider's limits and your own balance apply.
• Every action is visible, and Stop is immediate. Each tool call appears as it happens, the running cost of the session is shown in the header, and Stop aborts the request, ends the loop, and terminates any running command.
• Conversations are retained per folder, stored locally under the extension's own storage. Kiln: New Chat starts a fresh thread and Kiln: Chat History… lists previous ones.
• Kiln credits are available as an alternative to your own keys. When you are signed in, the model picker offers them alongside your providers. The distinction is shown where the choice is made: your own key goes directly to the provider and Kiln never sees the conversation, whereas a credits request is routed through Kiln's server.

Supported models

Anthropic Claude, OpenAI GPT, DeepSeek, and Moonshot Kimi are supported directly. Google Gemini, xAI Grok, Mistral, Meta Llama and Qwen are available through a single OpenRouter key. Each model is listed with its provider's published rate, and the session cost meter uses those rates.

Commands

All commands are available from the Command Palette under the Kiln category.

The status bar displays the current project, its version, and the number of modified files.

Configuration

Set kiln.api only when running a self-hosted instance:

// settings.json
"kiln.api": "https://your-instance/api/v2"

Data handling

API keys are held in the editor's secret storage. They are never written to disk in plain text, never recorded in logs, and never sent to Kiln.

In direct mode, conversations and file contents are sent only to the model provider whose key you supplied. They are stored locally and are not transmitted to Kiln.

In Kiln project mode, project contents and conversations are held by your Kiln account, as they are when using the web application.

Safeguards

• Pushes create versions; the history is preserved and reachable.
• Restoring an earlier version discards nothing — later versions remain.
• Pulling over unsaved work stops and reports how many files are affected. The option that preserves your work is always the default.
• Every refusal offers the route through it, and the overriding choice is never the default.
• A compiled React tree cannot be pushed, as its pages are generated from the static source and storing them back would overwrite the originals.

Support

Use Kiln: Report a Problem…, which assembles the relevant versions and logs and shows you the report before anything is sent.

Licence

Released under the MIT Licence.