Canvas

Canvas is an agent skill that gives coding agents a place to put their long-form output other than the chat transcript. Sessions can stay local or be published at an unguessable public URL backed entirely by Cloudflare.

When an agent produces something substantial — codebase research, an architecture explainer, an implementation plan, a comparison of options — it normally dumps hundreds of lines of markdown into the chat. The Canvas skill teaches the agent to route that output into a local browser workspace instead: it writes one HTML file per section, and a live workspace renders them.

What's in this repo

• skills/canvas-workspace/ — the skill itself: the instructions an agent reads to decide when and how to use Canvas. This is the product.
• bin/, server/, workspace/ — the llm-canvas CLI and the daemon that watches section files and serves the browser workspace. The skill drives these; bun run build bundles them into the skill so it can be installed standalone.

How it works

The contract is the filesystem. There is no API for the agent to learn:

1. The agent writes section files to .canvas/sessions/<slug>/sections/<id>.html.
2. The llm-canvas daemon watches that folder and serves a browser workspace.
3. When a file is written or edited, the affected section hot-reloads — no page refresh.

Because sessions are just plain HTML files under .canvas/ in your repo, they are readable, editable, and committable in any editor. The styled workspace view needs the daemon running; the files themselves do not.

It is also bidirectional: agents can emit choice and form widgets, users answer them in the browser, and the agent reads the responses back later.

Install

Canvas ships as an npm package that bundles a prebuilt CLI binary for your platform. You do not need Bun installed — the binary embeds its own runtime.

npm i -g llm-canvas
llm-canvas install-skill   # symlinks the skill into ~/.claude/skills/canvas-workspace
llm-canvas doctor          # verify the setup
llm-canvas login           # sign in with GitHub and approve this machine

macOS only — first-run Gatekeeper warning. The binary is not yet Apple-signed, so the first run is blocked by Gatekeeper. Clear the quarantine flag once and you are done:

xattr -dr com.apple.quarantine "$(which llm-canvas)"

Once installed, the agent starts the daemon itself when a session needs it — you do not run it by hand.

From source (contributors)

If you want to hack on Canvas, you'll need Bun:

bun install
bun bin/llm-canvas.ts doctor    # run the CLI without installing it
bun run build                    # produce per-platform binaries under packages/

CLI commands

| Command | What it does | | ----------------------------- | ------------------------------------------------------------- | | llm-canvas start | Start the daemon for the current project | | llm-canvas status | Report whether a healthy daemon is serving the current project| | llm-canvas doctor | Check the environment is set up correctly | | llm-canvas inbox | Read user feedback left in the workspace | | llm-canvas login [URL] | Authorize this machine through a browser | | llm-canvas whoami | Show the active Canvas account | | llm-canvas logout | Revoke this machine's Canvas session | | llm-canvas publish --slug S | Upload a session and its assets; print its public URL | | llm-canvas install-skill | Symlink the canvas-workspace skill into ~/.claude/skills | | llm-canvas uninstall-skill | Remove the symlink (only if it still points at this install) |

An MCP adapter is planned but not yet implemented. For now the integration is the file-write contract plus the shipped skill.

Public Canvas

Put images and other files in .canvas/sessions/<slug>/assets/ and reference them from a section as ../assets/<name>. Publish with:

llm-canvas publish --slug <slug>

The command uploads HTML and assets to the private llm-canvas R2 bucket and prints the public share URL. Re-running it updates the same URL. Each canvas is owned by the GitHub account that approved the CLI login; other signed-in users cannot update it. The hosted app is https://llm-canvas.our.workers.dev; bearer credentials stay in the local Canvas config and are never embedded in published files.

Documentation

• Architecture
• Design decisions

License

MIT