From Claude Code or Cursor to a deployed app.

An agent optimises for “it works when I run it here”. That is the right goal while you build, and it leaves the same handful of things behind almost every time:

• A dev command, not a start command. npm run dev, next dev or vite on localhost:3000 or 5173.
• Secrets in a local file. .env or .env.local with real keys. Sometimes an .env.example listing what the app expects, sometimes nothing.
• A local database. A dev.db SQLite file, or a DATABASE_URL pointing at Postgres on your machine.
• A lockfile that is missing or stale. The agent edited package.json and never re-ran the install, or the lockfile was never written.
• A port or URL typed in. app.listen(3000), or http://localhost:3000 as the API base in the frontend.
• An adapter for someone else’s platform. Ask for “a SvelteKit app” after reading a Vercel tutorial and you may get @sveltejs/adapter-vercel.

Until a project’s first successful deploy, Canner reads the upload before building and writes a warning for each of these into the build log, prefixed [canner] ⚠. The middle column is that text, or the error the build stops on. Fix them before the first deploy and you skip a round trip.

The CLI is the shortest path, and Claude Code or Cursor can run it in their own terminal. No Git repository is needed: canner deploy packs the folder and sends it to the same build pipeline a push uses.

npm install -g @canner-ca/cli     # Node 20 or later
canner login                      # paste a cnr_ token from Identity → API keys
canner init                       # creates the project, writes canner.json
canner db init                    # only if the app uses Postgres
canner env set OPENAI_API_KEY=sk-... --sensitive
canner deploy --follow            # uploads, builds, streams the log

canner deploy always leaves out node_modules, .git, .next, .nuxt, .turbo, .vercel and .svelte-kit. It does not read .gitignore, so a local .env goes up with everything else unless you exclude it. Add a .cannerignore next to package.json:

Set the real values as environment variables instead. --follow exits with 0 when the app is live and 1 when the build fails, so an agent running the command sees the failure and can read the log. The compressed upload is capped at 100 MB on Starter and 500 MB on Live.

# .cannerignore  (canner deploy does not read .gitignore)
.env
.env.*
*.sqlite
coverage/

The Canner MCP server gives the agent tools instead of shell commands: canner_create_project, then canner_save_app with the files, then canner_get_deployment to read the build log. Deleting projects, databases, domains or tokens is not exposed, so an agent can’t do it by mistake.

• Prompt: “Create a Canner project for this app, save it with the Canner tools, set the environment variables from .env.example, and give me the live URL. If the build fails, read the log and fix it.”
• CLI or MCP? canner_save_app needs the complete contents of every file sent through the conversation, up to 2,000 files. That is fine for a small app and slow for a large one. For anything big, have the agent run the CLI.
• Tokens: if you already ran canner login, the local server reads that token and you can leave out CANNER_TOKEN. For a token pasted into a third-party tool, use the “MCP connector (deploy only)” type.

• Logs: canner logs --follow streams the running app’s output. The agent can do the same with the MCP tools for errors and logs.
• Migrations: put npx prisma migrate deploy (or your tool’s equivalent) in the release command under Settings → Build, not in the build command.
• Sleep: Starter apps sleep after 3 hours without a request and wake on the next one. Live keeps 2 projects always on and adds custom domains.
• Everything else for AI-built apps is on the Host AI-built apps page.