Canvasbin
io.github.stewartmckeePublished by GitHub account stewartmckeev1.0.0Updated Nov 3, 2026
Publish HTML and Markdown pages at shareable links.
Overview
Canvasbin's hosted MCP server (canvasbin.dev); OAuth sign-in required; Streamable HTTP.
No AI overview yet: these facts come from the registry entry and the server itself.
- Endpoint
- canvasbin.dev
- Last checked
Installation
In SourceWeft
- Open Canvasbin in the dashboard and add it to a workspace.
- Connect the required credentials. They are stored privately and never shown on public pages.
- Enable the server for the chats that should use its tools.
Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.
Use with an agent
Paste this prompt into Claude Code, Codex, Cursor or another agent — or into SourceWeft chat. The agent reads this server's install guide, shows you its publisher, endpoint and sign-in, and adds it to your client once you agree. Keys never go into a command.
Read https://sourceweft.com/mcp/io.github.stewartmckee%2Fcanvasbin/install.md and connect the MCP server it describes to my client. Before changing any config, show me its publisher, endpoint and sign-in method and wait for my OK. Never put an API key in a command or in a file I would commit; ask me where it should come from.Other MCP clients
Add this to your client's mcpServers config. The client completes OAuth sign-in on first connect.
{
"mcpServers": {
"canvasbin": {
"type": "http",
"url": "https://canvasbin.dev/mcp"
}
}
}README
Canvasbin
Pastebin for the agentic age.
Canvasbin was built for sharing what agents produce. An agent can design a page in HTML — a chart, a dashboard, a report, or an entire Vue application — and Canvasbin publishes it. HTML or Markdown in, a link out. No deploy, no CMS. Agents publish; humans share.
Use the hosted service at canvasbin.dev, where you sign in once and your pages live in a library you can share, or run your own copy with accounts off and no sign-in at all.
The link is the permission. Share the read link. Keep the write link. Anyone with the write link can edit or delete the page, including the agent that created it. Ask again and the same URL updates in place.
create_page returns the write key to the agent, so that key is in the chat
history. Share html_url from the transcript. The write key stays there so
the agent can edit the page later; treat the transcript like the editor.
What you get
The address itself picks the view:
Pages can be a fragment or a full HTML document. Markdown is rendered for you. Updates are live on the same URLs, so an agent can revise a page without minting a new link.
Agent scratchpads (agentic_only)
Agents can create agentic-only pages: external scratchpads for context shared
across MCP clients or backups of what matters from a chat. Set agentic_only: true on create_page or update_page. They stay out of the web library unless
the user turns on “Show agent scratchpads”; list_pages always includes them
with the flag set. Passing team with agentic_only: true stores a personal
page, not a team page. See
Share context across platforms
and Backup AI chats.
Hand it to an agent
Point Claude Code, Cursor, Codex, or any MCP client at
https://{your-canvas-bin-host}/mcp. There is no API key: a self-hosted copy
with accounts off needs no sign-in, and canvasbin.dev asks you to sign in once
through the client. In Cursor, install the Canvasbin plugin instead. Ask for a
page:
Build a dashboard of this week's numbers and publish it on Canvasbin. Give me the link.
You get html_url to forward. The agent also receives write_url and
write_key, and those stay in the chat history for the next edit. Setup for
each client is below.
Examples use https://{your-canvas-bin-host}. Replace that with the origin
where you run Canvasbin. On your own machine that is http://localhost:8080.
Run it
Self-hosting is the default. Accounts stay off: there is no sign-in, and a
page is created with its read and write links. You need Docker. Postgres
stores the pages. CircleCI publishes the app image on each push to main
as stewartmckee/canvasbin
(linux/amd64). It includes the editor and headless Chromium for PDF
export. Each build also tags 0.1.<build number>.
ACCOUNTS defaults to off. Leave it that way for a private bin.
Docker Compose
This is the sample to run. The same file is
examples/compose.yaml. It is not
docker-compose.yml, which is only for developing this repository.
Open http://localhost:8080.
Docker Run
The same two containers without a compose file. restart brings the app
back up if Postgres is not ready on the first start.
Open http://localhost:8080.
For anything other than local use, put a TLS reverse proxy in front and set
PUBLIC_BASE_URL to that https origin with no trailing slash. Do not
publish port 8080 on a public interface; only the proxy should reach the
app. The proxy must append the caller to X-Forwarded-For. Canvasbin uses
the rightmost address, and only when the TCP peer is a private or loopback
address.
Pin 0.1.<build number> instead of latest when you want upgrades to be deliberate.
To build the image from this checkout instead of pulling it:
Use that name in the docker run above. docker compose up --build in the
repository root is the development stack, not this sample.
Develop from source
Requirements: Go 1.27, Node.js 24, Docker.
API (proxied by the Vite dev server):
UI:
Open http://localhost:5173. PDF export looks for Chrome or Chromium on
PATH, or the binary in CHROME_BIN. The published image sets
CHROME_BIN=/usr/bin/chromium-browser.
Configuration
Copy .env.example for a local env file. Never commit real secrets.
Page content is limited to 1 MiB. Creating pages is rate limited (60 per 10
minutes per client IP). PDF renders are limited too. Behind a proxy, append
the caller to X-Forwarded-For ($proxy_add_x_forwarded_for). Canvasbin
uses the rightmost address in that header, and only when the connection
comes from a private or loopback address, so a client cannot choose a fresh
rate-limit bucket by sending its own value. Do not publish the app port on
a public interface; the proxy should be the only thing that can reach it.
Accounts (optional)
With ACCOUNTS=off (the default), Canvasbin stays authless: create a page, share
the read link, keep the write link.
With ACCOUNTS=on, people can register with email and password. Mailgun sends
the verification link, so MAILGUN_SENDING_KEY and PUBLIC_BASE_URL are
required. Signed-in users get a library and can claim a write-key page.
Creating a page then requires signing in; existing pages keep working by
their read and write links without an account. The Docker samples leave
accounts off.
Connect an agent
Canvasbin speaks MCP over Streamable HTTP. The write key from create_page is
what lets an agent edit or delete that page later.
The endpoint is https://{your-canvas-bin-host}/mcp. On your own machine
that is http://localhost:8080/mcp. With accounts off, no API key or login is
required. With accounts on, every MCP request needs a signed-in account
(OAuth), so clients prompt for sign-in when they connect. Page links in the
browser keep working by key without signing in.
Sign in from Claude, ChatGPT or Cursor
Point the connector at https://canvasbin.dev/mcp (or your hosted origin).
The first request returns HTTP 401 with a WWW-Authenticate challenge.
The client opens Canvasbin’s OAuth consent screen; after you approve, it
retries with a bearer token. Discovery documents live at
/.well-known/oauth-protected-resource/mcp and
/.well-known/oauth-authorization-server.
Cursor plugin
This repository is also a Cursor plugin (.cursor-plugin/plugin.json). The
Marketplace install always connects Cursor to https://canvasbin.dev/mcp
and adds a canvasbin skill that tells the agent when to publish a page and
which link to hand back. Sign-in is OAuth on that host: Cursor opens the
Canvasbin consent screen the first time, and the plugin never sees the
account password. To use your own host instead, configure the server by hand
as shown under Other clients.
Claude Code
Add it for every project on this machine:
Or only the project you are in (the default scope):
Check it with claude mcp list, then /mcp inside a session. Ask Claude to
publish a page. A typical request:
Create a Canvasbin page that explains this change, and give me the link to share.
create_page returns four links, plus the keys. html_url is the one to
share. write_url and write_key come back to Claude and remain in the
chat history, which is what lets a later turn call update_page without
opening the browser.
To share the server with everyone who clones a repository, commit a
.mcp.json in that repository:
Claude Code asks before it uses a project .mcp.json. "type": "http" is
required. streamable-http is accepted as an alias. This repository commits
that file at the root, pointed at https://canvasbin.dev/mcp, so
cursor.directory can find the server.
Other clients
No headers are required.
Cursor. In ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one
project):
Codex.
That writes this into ~/.codex/config.toml:
Claude Desktop. Add a custom connector and set the URL to
https://{your-canvas-bin-host}/mcp. Desktop does not take a remote URL in
claude_desktop_config.json.
VS Code. In .vscode/mcp.json:
Tools
create_page— publish HTML or Markdown. Returnshtml_url,pdf_url,read_url,write_url, and the keys. When publishing into a team, the agent is asked to read that team's style guide first.get_page— read a page by read key, write key, or URL.update_page— replace content, title, or format. Needs the write key. Omitted fields stay as they are.contentreplaces the whole page.delete_page— delete a page. Needs the write key.get_style_guide/set_style_guide— read, create, or update a team's style guide. These are registered only when accounts are on. Any member of that team can create the guide or replace it later.create_pagetells the agent to callget_style_guidebefore writing a team page.
Tools that take a key accept the bare key or any Canvasbin URL. HTML may be a fragment or a full document. Scripts in the rendered page run in a sandbox.
Issues
Use the issue templates so reports are actionable:
- Bug report — what you expected, what happened, how to reproduce it, and the image tag or commit.
- Feature request — the job you are trying to do, and the behaviour you want.
Search existing issues before opening a new one. Security reports go through private advisories, not a public issue. See SECURITY.md and CONTRIBUTING.md.
License
MIT.
Source: README.md at commit 37cc879
Tools
0Version history
1- v1.0.0LatestNov 3, 2026
Score
85% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 daysRequired25 pts
- Auth declaredRequired10 pts
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts