Deploy to Cloud Engine

What This Is

A cloud engine is a user-owned slice of Internet Computer capacity, administered from a web console (by default https://opencloud.org). Each engine runs on a single subnet. This skill takes a project that already builds and gets it deployed onto that engine, from a coding agent.

This skill only covers the cloud-engine-specific steps: linking the CLI to the engine’s console identity, and a subnet-targeted deploy. For everything else about the CLI (icp.yaml, recipes, environments, bindings, identities), load the icp-cli skill.

Before running any icp command you are unsure of, run icp <subcommand> --help (e.g. icp identity link --help, icp deploy --help) to confirm the command and flags exist. Do not infer flags. Authoritative reference: https://cli.internetcomputer.org/llms.txt

What You Need

Two values. Look for them first in icp.yaml or earlier in the conversation. One has a default; the other you must ask for:

Console origin — the URL the user signs in to their cloud engine console with. Defaults to https://opencloud.org (the main OpenCloud console). It is used as the --auth origin in Step 1 so the linked CLI identity derives the same principal that administers the engine. Use the default, but say so and give the user a chance to override before linking:
- Say: “I’ll link the CLI against https://opencloud.org, the default console. If you sign in to your engine console at a different URL, tell me now.”
- Only use a different origin when the user names one — never substitute another URL on your own; the --auth origin determines the derived principal (see Pitfall 2).
Subnet id — the subnet the engine deploys to, required by icp deploy --subnet. There is no default; never guess it. The user finds it on the engine’s App Center / Applications page in the console. If absent, ask and do not proceed without it:
- Ask: “What is your engine’s subnet id? It is shown on your engine’s App Center / Applications page.”

Record both so you do not re-ask within the session.

Prerequisites

icp on $PATH — see the icp-cli skill to install. Verify with icp --version (this skill’s commands are verified against 0.3.0 and 1.0.2). If the installed version differs, confirm the flag set with icp <cmd> --help before running — flags have changed across major versions.
A project that already builds. If it does not build or package yet, set that up first (see the icp-cli skill), then return here.
macOS: icp stores its data under ~/Library/Application Support/org.dfinity.icp-cli/. If the shell cannot write there (Operation not permitted from macOS TCC, e.g. when commands run through a bridge), redirect the data to an unprotected path for the session — HOME=/tmp/icp-home icp … — and keep that same HOME on every subsequent icp command, or the later commands will not see the linked identity.

Step 1 — Link the CLI to your engine identity (once per machine)

The CLI must sign as the same identity that administers the engine — that is the principal you log in to the console with.

Step 1.0 — First determine WHERE the CLI runs relative to the browser

icp identity link web completes the sign-in via a redirect to http://127.0.0.1:<port> on the machine where the CLI runs. The browser in which the user completes the Internet Identity sign-in must be able to reach that loopback address. Determine the environment before linking:

CLI and browser on the same machine, with network (local development, or a terminal-integrated agent on the user’s own machine) → use the normal link flow below.
CLI in a remote or isolated sandbox, with network (the agent runs in the cloud; the user’s browser is on a different machine) → the normal flow cannot work: the sign-in URL carries a callback=http://127.0.0.1:<port> bound to the sandbox, the user’s browser resolves 127.0.0.1 to its own machine, and the delegation never reaches the CLI. Later commands then fail with authorization errors. Use the delegation handoff below instead.
CLI shell with no network at all (e.g. a sandboxed device bridge on the user’s own machine: DNS blocked, HTTPS requests fail, writes limited to /tmp) → no icp network command can run there — not the link, and not icp deploy either. The delegation handoff does not help: it only moves signing authority, not network access. Do not retry, tunnel, or proxy. Prepare the project and hand the user one script to run in their real terminal — see “No-network CLI host” below.

When in doubt, probe before linking: curl -sI https://<console-origin> failing (or DNS not resolving) from the CLI shell means the no-network case.

Never present a sandbox 127.0.0.1 URL to the user as something to open in their browser — it is unreachable from their machine. And do not invent a headless flag on icp identity link web: as of icp-cli 1.0.2 it has none (confirm with icp identity link web --help).

Delegation handoff (CLI host and browser on different machines)

icp identity delegation transfers signing authority from an identity linked on the user’s machine to a session key on the CLI host, without the browser ever reaching the CLI host. It requires icp installed on the user’s machine too, and exists in icp-cli 1.0.x — verify with icp identity delegation --help on both machines. If the user’s machine has icp and you can run commands there, the simplest path is to run this whole skill there instead; otherwise:

On the CLI host (sandbox) — create a pending identity with a fresh session key:
```
icp identity delegation request <your-identity-name>
```
It prints the session public key as a PEM to stdout. Give that PEM block to the user. (If the default --storage keyring fails in a headless sandbox, retry with --storage plaintext.)
On the user’s machine — the user links there if not already linked (icp identity link web <local-name> --auth <console-origin> — the normal flow works locally), saves the PEM to a file, and signs a delegation to it:
```
icp identity delegation sign --identity <local-name> --key-pem session-key.pem --duration 8h > delegation.json
```
--duration takes e.g. 30m, 8h, 1d; optionally --canisters <ids> restricts the delegation. The user sends delegation.json back. If sign fails with delegation for identity <name> has expired or will expire within 5 minutes, the local web session has lapsed — run icp identity reauth <local-name> (the normal browser flow, which works locally) and retry.

On the CLI host — complete the pending identity and activate it:

icp identity delegation use <your-identity-name> --from-json delegation.json
icp identity default <your-identity-name>
icp identity principal   # must print the user's console principal

Treat delegation.json as a time-limited credential: whoever holds the sandbox’s session key can sign as the user’s console principal until it expires. Keep the duration short and re-run the handoff when it lapses.

No-network CLI host — hand the user one script

If the CLI shell cannot reach the network, the user’s real terminal is the deploy machine — and because their terminal and browser share a loopback there, the normal link flow works; no delegation handoff is needed. Your job shifts to preparation:

Get the built project into a directory the user’s terminal can reach.

Write one script at the project root that does the whole network-bound sequence — adapt this template (fill in the real identity name, console origin, and subnet id; do not leave placeholders):

#!/usr/bin/env bash
set -euo pipefail
icp --version
icp identity list | grep -E '^\*? *<your-identity-name> ' >/dev/null || icp identity link web <your-identity-name> --auth <console-origin>
icp identity default <your-identity-name>
icp deploy -e ic --subnet <subnet-id>

Ask the user to run it in their terminal, complete the Internet Identity sign-in when the browser opens, and paste the output back.
Verify from the pasted output (canister ids, frontend URL) and continue — a re-deploy for the Step 2 __META_* variables goes through the same script.

First check what already exists:

icp identity list      # names + principals; * marks the active identity

The list does not show which console (if any) an identity was linked against — that cannot be determined from the CLI. Decide like this:

Only anonymous (or plain local identities) listed — no web-linked identity exists; run the link command below.
An identity the user recognizes as their engine identity (by name or principal) — set it active (below) and skip to Step 2.
Unsure — ask the user, or simply relink under a new name; linking again is cheap and safe.

To link, run this, substituting a name the user picks — <your-identity-name> is any local label, not a fixed value (do not hardcode something like my-engine-admin); reuse the same name in every command below:

icp identity link web <your-identity-name> --auth <console-origin>

Use https://opencloud.org as <console-origin> unless the user named a different console. Never omit --auth: the flag has a built-in default (https://id.ai) that is not your console and silently derives the wrong principal.
The command first waits at a “Press Enter to log in” prompt before anything happens. Run it interactively when you can; in a non-interactive shell (e.g. a background process) pipe a real newline:
```
printf '\n' | icp identity link web <your-identity-name> --auth <console-origin>
```
Do not redirect stdin from /dev/null — the bare EOF does not satisfy the prompt, and the command sits on “Press Enter to log in” indefinitely with no browser ever opening.
After Enter it opens a browser tab. The user completes the Internet Identity sign-in there. Wait for them to confirm before continuing — you cannot complete the sign-in for them.
--auth must be the exact console origin (scheme + host), e.g. https://opencloud.org. A mismatched origin derives a different principal, and the engine will reject the deploy as unauthorized.
This is a one-time, per-machine step.

Then make it the active identity and verify:

icp identity default <your-identity-name>
icp identity default     # prints the active identity name
icp identity principal   # prints the principal the deploy will sign as

Step 2 — Name the app (and give it an icon) in the console (recommended)

By default, CLI-deployed canisters appear on the engine console’s Applications page as bare rows labelled only by their principal id. A set of canister environment variables makes the console group them into a single named application with readable per-canister labels, an “Open” button, and an icon. Set them once in your project config:

__META_PROJECT — the application name. Canisters that share the same value are grouped into one named app, so set an identical value on every canister of the app.
__META_NAME — the per-canister display label (e.g. Backend, Frontend).
__META_MAIN_CANISTER — the literal string "true" on exactly one canister (the entry point, usually the frontend/asset canister). This marks the app’s main canister: the console reads __META_BASE_URL and __META_ICON_PATH only from it, and the “Open” button targets it.
__META_BASE_URL — an absolute https:// URL, set on the main canister (e.g. the frontend canister’s URL https://<frontend-canister-id>.icp.net, or a custom domain). When present and valid, it is the URL the “Open” button opens; when absent or not https, the “Open” button falls back to the main canister’s gateway URL. It is also the base that __META_ICON_PATH resolves against.
__META_ICON_PATH — the path to the app icon, resolved against __META_BASE_URL to form the icon the console renders (e.g. /favicon.svg → https://<base>/favicon.svg). Set it on the main canister, alongside __META_BASE_URL.

The icon and “Open” link are read only from the main canister (the one marked __META_MAIN_CANISTER: "true") — __META_BASE_URL / __META_ICON_PATH on any other canister are ignored.

Set them under each canister’s settings.environment_variables — this is valid alongside a recipe. With per-canister canister.yaml files:

# frontend/canister.yaml
name: frontend
recipe:
  type: "@dfinity/static-site@v0.3.3"
  configuration:
    build:
      - npm install
      - npm run build
    dir: dist
settings:
  environment_variables:
    __META_PROJECT: "My App"
    __META_NAME: "Frontend"
    __META_MAIN_CANISTER: "true"
    __META_BASE_URL: "https://<frontend-canister-id>.icp.net"
    __META_ICON_PATH: "/favicon.svg"

# backend/canister.yaml
name: backend
recipe:
  type: "@dfinity/motoko@v5.0.0"   # v5 reads main/candid from mops.toml ([canisters.backend]) — see the icp-cli skill
settings:
  environment_variables:
    __META_PROJECT: "My App"
    __META_NAME: "Backend"

For a single inline icp.yaml (canisters defined there directly), put the same settings.environment_variables block under each canister entry. Note the inline form: canisters is an array of {name, recipe, settings} items, not a map keyed by canister name:

# icp.yaml — canisters defined inline
canisters:
  - name: frontend
    recipe: # … as in the canister.yaml example above
    settings:
      environment_variables:
        __META_PROJECT: "My App"
        __META_NAME: "Frontend"
        __META_MAIN_CANISTER: "true"
        __META_BASE_URL: "https://<frontend-canister-id>.icp.net"
        __META_ICON_PATH: "/favicon.svg"
  - name: backend
    recipe: # … as in the canister.yaml example above
    settings:
      environment_variables:
        __META_PROJECT: "My App"
        __META_NAME: "Backend"

Notes:

icp-cli merges these with the PUBLIC_CANISTER_ID:<name> variables it injects automatically at deploy time — the asset canister keeps serving and the app keeps working. (Verified against icp-cli 0.3.0.)
All values are strings; __META_MAIN_CANISTER must be the exact string "true".
They are applied during icp deploy (the “Setting environment variables” step). After deploy, confirm with icp canister settings show <name> -e ic.

Icon specifics (the console builds the icon as __META_BASE_URL + __META_ICON_PATH):

Both must be present and on the main canister for an icon to appear — there is no fallback. __META_ICON_PATH alone does nothing.
__META_BASE_URL must parse as an absolute https:// URL. A bare host, an http:// URL, or a data: / javascript: value is rejected: the icon then does not render, and the “Open” button falls back to the main canister’s gateway URL (it does not disappear). (The console validates the scheme before using it.)
__META_ICON_PATH is a path to an asset your frontend actually serves (e.g. /favicon.svg), not an inline image. The resolved URL is rendered as an <img> src, so it must return an image. Do not put a data: URI here: engine env values are length-capped (≤128 chars observed), so it would not fit, and the field is a path by design.
The frontend canister’s id is only known after the first deploy. The usual flow is: deploy once, read the frontend canister id from the output, set __META_BASE_URL to https://<that-id>.icp.net (and __META_ICON_PATH), then re-deploy to apply. If you control a custom domain for the app, you can set it up front instead.

Pin the Internet Identity derivation origin (if the app signs users in)

Internet Identity derives a principal per origin. An app reached at both its canister address and a custom domain signs the same person in as two different users.

The engine console names the app’s canister address as the origin to derive from. It cannot be renamed or removed, so it survives adding, changing, or dropping a custom domain. It is the address of the canister that serves your frontend (the asset/static-site canister the browser loads the app from), built from that canister’s id:

https://<frontend-canister-id>.icp.net

That is usually the canister you marked __META_MAIN_CANISTER: "true", but that flag is a console display setting, not a guarantee. If you marked a backend canister, still derive from the frontend canister. A backend canister cannot be a derivation origin: the browser is never on it, and it cannot serve the .well-known/ii-alternative-origins file Internet Identity fetches from that origin to validate the claim.

Build the value from the canister id, not from __META_BASE_URL. The two match only while __META_BASE_URL still points at the canister address. Pointing it at a custom domain is normal and supported (see the icon notes above), and the two then differ. A custom domain is the one value that must never become the derivation origin.

Do this before the app has users, not after. A domain that has already collected sign-ins cannot be repointed at a derivation origin without orphaning every account created under it. If the app might ever get a custom domain, set derivationOrigin on the first deploy even while the canister address is the only origin — it costs nothing then and cannot be applied retroactively without losing accounts.

For the mechanics — the derivationOrigin option and the .well-known/ii-alternative-origins file, including the _headers entry the @dfinity/static-site recipe needs — load the internet-identity skill.

Caffeine apps are the exception: Caffeine owns their address, so it is not the app’s to derive from. The console withholds the canister-address row for them.

Version metadata (recommended)

Embed build provenance into each canister’s wasm so tooling and the console can show what version is running (console support is rolling out; the metadata is already readable — see the verify note below). Add this by default on every deploy — do not wait for the user to ask. All official recipes (motoko, rust, static-site, asset-canister, prebuilt) accept a metadata list under the recipe configuration; each entry is baked into the wasm as a custom section. Values are interpolated into a shell command at build time, so $(…) command substitution works. (Verified by live builds against @dfinity/motoko v4.1.0 and v5.0.0 on icp-cli 1.0.2.)

Git project (check first: git rev-parse HEAD succeeds):

recipe:
  type: "@dfinity/motoko@v5.0.0"
  configuration:
    metadata:
      - name: service:git:sha
        value: $(git rev-parse HEAD)$(git diff --quiet HEAD 2>/dev/null || echo +dirty)
      - name: service:git:origin
        value: $(git remote get-url origin)
      - name: service:git:updated_at
        value: $(git log -1 --format=%cI)
      - name: service:version
        value: $(node -p "require('./package.json').version" 2>/dev/null || echo 1.0.0)

Non-git project — the git substitutions above do not fail the build; they silently bake garbage (service:git:sha becomes the literal +dirty, service:git:origin comes out empty). Use only an explicit version, and bump it on each subsequent deploy:

metadata:
  - name: service:version
    value: "1.0.0"

Notes:

Use these exact names (service:git:sha, service:git:origin, service:git:updated_at, service:version) — they are the agreed convention the console will read; differently named entries will not be picked up.
Prefer command results over literals where possible: +dirty on the sha flags uncommitted changes, so a deploy is traceable to its exact source state.
Values must be deterministic for a given source tree: sha, origin, the last commit’s date, a version string. Never embed a build time ($(date)) or other build-time-varying values. The commit date (%cI) is allowed because it is a property of the commit, not of the build — but note it means “code last changed”, not “last deployed”: deploying an old commit shows the old date. The deploy time itself comes from the canister history the network records automatically.
If the project has no package.json (or no version field), the fallback 1.0.0 applies; replace it with the project’s real versioning scheme when one exists.
The sections are injected as icp:private, readable by the canister’s controllers (you, and the console acting as you) via read_state — no canister call needed. Verify after deploy, with the linked identity active: icp canister metadata <canister-name> service:git:sha -e ic.

Step 3 — Deploy to the engine’s subnet

From the project root:

icp deploy -e ic --subnet <subnet-id>

-e ic targets mainnet (the engine runs on an IC subnet); --subnet <subnet-id> pins the deploy to your engine’s subnet. Confirm the exact flags with icp deploy --help before running if unsure.
Deploying consumes capacity on the engine; make sure the engine has room.

Alternative — packaged upload. If the project is distributed as a built .icp package and a direct icp deploy is not available, upload the bundle on the console’s App Center via “Upload a custom app” instead.

Step 4 — Verify

The icp deploy output reports the deployed canister ids.
The canisters appear on the engine’s Applications page in the console; each canister’s detail view offers an “Open in browser” link.
If you set the metadata in Step 2, the canisters are grouped under your __META_PROJECT name with their __META_NAME labels, and the main canister shows an “Open” button — instead of bare principal rows. With __META_BASE_URL + __META_ICON_PATH set, the app also shows its icon (allow for a short console cache delay).
A frontend (asset) canister is served at https://<frontend-canister-id>.icp.net.

Report the deployed canister ids (and the frontend URL, if any) back to the user.

Code that runs on the engine

A cloud engine runs on a CloudEngine subnet with protocol-level call rules that differ from normal Application subnets: engine canisters hold 0 cycles and must never attach cycles to any call, cross-subnet calls must be bounded-wait, HTTPS outcalls are made directly and free, and cycle-bearing cross-subnet targets (exchange-rate XRC, threshold ECDSA/Schnorr, vetKD) go through the engine’s funded console proxy canister. Before writing or debugging any canister code that will run on the engine, load the cloud-engine-canisters skill — it has the rules, the proxy interface, and the pitfalls.

Common Pitfalls

Sign-in not completed. Running icp identity link web … but not finishing the Internet Identity sign-in in the browser leaves the CLI unlinked; later commands fail with authorization errors. Re-run and wait for the user to confirm the browser flow finished. If no browser ever opened, the command is stalled at the “Press Enter to log in” prompt — relaunch with a piped newline, printf '\n' | icp identity link web …, never < /dev/null (see Step 1). If the CLI runs in a remote sandbox, re-running can never complete — see Pitfall 10 and the delegation handoff in Step 1.0.
Wrong --auth origin. Using any URL other than the console origin the user signs in with derives a different principal, and the engine rejects the deploy as not authorized. Relink with the exact console URL. If the deploy is rejected as unauthorized after linking against the default https://opencloud.org, ask the user for the exact URL they sign in with and relink.
Guessing the subnet id. Never invent it — the deploy fails or targets the wrong subnet. It is on the engine’s App Center / Applications page; ask the user.
Deploying with the anonymous identity. The default local identity is anonymous and is not the engine admin. You must link and icp identity default <your-identity-name> first.
Using dfx. This ecosystem uses icp, never dfx. The correct sequence is icp identity link web <name> --auth <console-origin> (Step 1), then icp deploy -e ic --subnet <subnet-id> (Step 3). See the icp-cli skill.
Skipping the app metadata. Without __META_PROJECT (Step 2), the canisters still deploy and work but render as bare, unnamed principal rows in the console. Setting __META_* is what produces a named app with labelled canisters and an “Open” button.
Wrong __META_MAIN_CANISTER value. It is matched as the exact string "true". A boolean, "True", or marking more than one canister means no (or the wrong) “Open” button. Mark exactly one entry-point canister.
Inventing an icon variable. The icon variable is __META_ICON_PATH (a path resolved against __META_BASE_URL). Do not guess __META_ICON, __META_LOGO, or __META_ICON_LINK — they are ignored, so the icon silently never appears.
Icon set on the wrong canister, or without a base URL. The icon is read only from the main canister and needs both __META_BASE_URL (a valid absolute https:// URL) and __META_ICON_PATH. Setting the icon path on a side canister, omitting the base URL, or giving a non-https / data: base means no icon renders. (The “Open” button still works — it falls back to the main canister’s gateway URL — so a bad base URL costs the icon and the custom Open URL, not the button.)
Assuming the browser and CLI share localhost. icp identity link web returns the delegation to 127.0.0.1:<port> on the CLI host. If the CLI runs in a remote sandbox while the user’s browser is on a different machine, the sign-in never completes and later commands fail with authorization errors — no amount of re-running the link fixes it. See Step 1.0: use the delegation handoff (icp identity delegation request / sign / use), or run the link and deploy where the browser and CLI share a loopback.
Expecting the delegation handoff to fix a missing network. The handoff moves signing authority, not connectivity — icp deploy still needs the network from the shell that runs it. If the CLI shell has no network at all (a sandboxed device bridge: DNS blocked, HTTPS fails), no icp network command can run there, link or deploy. Do not offer to “do the rest” from that shell and do not tunnel — hand the user one script for their real terminal (see “No-network CLI host” in Step 1.0), where the normal link flow works because terminal and browser share a loopback.
Build timestamps in wasm metadata. Metadata is baked into the wasm and must be deterministic — a build time ($(date)) changes every build even with identical source, breaking reproducibility and changing the module hash on every deploy. The deterministic alternative is the last commit’s date, service:git:updated_at = $(git log -1 --format=%cI) — a property of the source tree, not the build. The deploy time itself comes from the canister history recorded by the network, never from metadata.
Git metadata substitutions in a non-git project. Outside a git repository, $(git rev-parse HEAD) does not fail the build — it silently bakes garbage: service:git:sha becomes the literal +dirty and service:git:origin comes out empty. Check for a git repo first (git rev-parse HEAD succeeds); if there is none, set only service:version with an explicit value (or git init and commit before deploying, if version control is wanted anyway).
Letting an engine app collect sign-ins before pinning a derivation origin. Internet Identity principals are per-origin, so adding a custom domain later turns every existing user into a stranger at the new address — and the fix cannot be applied retroactively without orphaning the accounts already made under the old origin. On the first deploy of any app that uses II, set derivationOrigin to the address of the canister that serves the frontend, built from its canister id (https://<frontend-canister-id>.icp.net), even when that is currently the app’s only origin. Do not copy __META_BASE_URL: it is allowed to point at a custom domain, and a custom domain is exactly what must not become the derivation origin. See the internet-identity skill for the .well-known/ii-alternative-origins half.

If a referenced skill is not already available, install it the same way this one was installed — npx skills add dfinity/icskills --skill <name> — or read it at https://skills.internetcomputer.org/skills/<name>/.

cloud-engine-canisters — the call rules for code that runs on the engine: never attach cycles, bounded-wait cross-subnet calls, direct HTTPS outcalls, and the console proxy for cycle-bearing cross-subnet targets (XRC, threshold signing, vetKD). Load it before writing or debugging engine canister code.
icp-cli — general icp CLI usage (icp.yaml, recipes, environments, bindings, identities). Load it for anything beyond this cloud-engine deploy flow — in particular when the project does not build or package yet.
internet-identity — details of the Internet Identity sign-in that Step 1 triggers in the browser.