Canvas Assistant

Canvas Assistant is an optional app that brings a chat assistant into the workspace, able to look at the files you point it at and propose changes you approve. It is developed separately from the desktop, and you choose the model endpoint it talks to.

Prerequisites

• ShellCanvas 0.1.13 or newer on Windows, Linux or macOS. The same platform-neutral Assistant app package is used everywhere.
• An endpoint that speaks the OpenAI Responses API or a compatible Chat Completions API, and its API key. A local model server works too.
• For the host tools, a connected workspace. For plain conversation, none.

The assistant is not bundled: you install it like any other app, from its public repository at github.com/techartdev/ShellCanvas-Assistant.

Steps

Install it

1. Open App Manager → Installed → Recommended by ShellCanvas and choose Review on Canvas Assistant. You can also use Add apps → From GitHub with techartdev/ShellCanvas-Assistant and a branch, tag or commit.
2. Review the version and requested access, then choose Install app. A recommendation does not approve permissions or install anything by itself.

It asks for a broad set of permissions, because it is a broad tool: its own storage, system dialogs, the network permission for your model endpoint, clipboard access, file read, edit and create, and console access. Every one of them is listed in the review, and you can decline by not installing.

Connect a model

Open the assistant and configure its connection. ShellCanvas — not the app — shows the dialog that asks for the endpoint URL and API key, and stores the key in your operating system's credential store, or keeps it for the session only.

The app never receives the key. It sends requests through the desktop to exactly the endpoint you configured, with no redirects and no other destination. A plain http:// endpoint is allowed for a model server on your own machine, with a warning.

Work with it

• Ask questions, with streaming answers you can cancel or retry.
• Attach context deliberately: selected text, images, and files you choose from the workspace.
• Let it use host tools: discovering the workspace, selecting and reading files, proposing file changes and console commands.

Every change it proposes — writing a file, running something in a console — waits for your explicit approval. Conversation history is kept locally on your computer.

Expected result

An assistant that understands which host you are working on, can read the files you show it, and never touches the machine without your approval. Terminal and Files keep working exactly as before; nothing depends on the assistant being installed.

Limitations

• You bring the model. There is no bundled provider, no account and no OAuth. Whatever you send goes to the endpoint you configured, under that provider's terms.
• Native app-frame probes passed on Windows, Linux and Apple silicon. Live Assistant sessions on Linux and macOS still need manual testing.
• Its permissions are wide by design. They are the same permissions any app would need for those features, enforced by the same broker, but an assistant that can read files and run commands deserves a deliberate decision.
• Keys belong to the installation. Removing and reinstalling the app loses saved connections, and you configure the endpoint again.
• It is a separate project, with its own release cadence and its own issue tracker.

Related guides

• Install and update apps — the installation flow in detail.
• Extension permissions and trust — what each grant allows.
• Network and credentials API — how the key is kept away from the app.
• App installation and permissions — when something is refused.