ICP CLI

What This Is

The icp command-line tool builds and deploys applications on the Internet Computer. It replaces the legacy dfx tool with YAML configuration, a recipe system for reusable build templates, and an environment model that separates deployment targets from network connections. Never use dfx — always use icp.

Before generating any icp command not explicitly documented here, run icp --help or icp <subcommand> --help to verify the command and its flags exist. Do not infer flags from dfx equivalents — the CLIs are not flag-compatible.

Installation

npm install -g @icp-sdk/icp-cli @icp-sdk/ic-wasm

ic-wasm is required when using official recipes (@dfinity/rust, @dfinity/motoko, @dfinity/asset-canister) — they depend on it for optimization and metadata embedding. Requires Node.js >= 22. Also available via Homebrew and shell script installer — see the icp-cli releases.

Linux note: On minimal installs, you may need system libraries: sudo apt-get install -y libdbus-1-3 libssl3 ca-certificates (Ubuntu/Debian) or sudo dnf install -y dbus-libs openssl ca-certificates (Fedora/RHEL).

Prerequisites

For Rust canisters: rustup target add wasm32-unknown-unknown
For Motoko canisters: npm i -g ic-mops and a mops.toml at the project root with the Motoko compiler version and a [canisters] entry:
```
[toolchain]
moc = "1.9.0"

[canisters.backend]
main = "src/backend/main.mo"
```
The @dfinity/motoko@v5+ recipe compiles via mops build <canister-name>. The canister name in icp.yaml must exactly match a key in [canisters] — a missing or mismatched key causes mops build to fail with No Motoko canisters found in mops.toml configuration (see Pitfall 17). Without mops.toml, the recipe fails because mops is not found. Templates include mops.toml automatically; for manual projects, create it before running icp build. Load mops-cli for [canisters] configuration options, dependency management, and mops build details.

Common Pitfalls

Using dfx instead of icp. The dfx tool is legacy. All commands have icp equivalents — see references/dfx-migration.md for the full command mapping. Never generate dfx commands or reference dfx documentation. Configuration uses icp.yaml, not dfx.json — and the structure differs: canisters are an array of objects, not a keyed object.
Using --network ic to deploy to mainnet. icp-cli uses environments, not direct network targeting. The correct flag is -e ic (short for --environment ic).
```
# Wrong
icp deploy --network ic
# Correct
icp deploy -e ic
```
Note: -n / --network targets a network directly and works with canister IDs (principals). Use -e / --environment when referencing canisters by name. For token and cycles operations, use -n since they don’t reference project canisters.
Using a recipe without a version pin. icp-cli rejects unpinned recipe references. Always include an explicit version. Official recipes are hosted at dfinity/icp-cli-recipes.
```
# Wrong — rejected by icp-cli
recipe:
  type: "@dfinity/rust"

# Correct — pinned version
recipe:
  type: "@dfinity/rust@v3.3.0"
```
Writing manual build steps when a recipe exists. Official recipes handle Rust, Motoko, and asset canister builds. Use recipe: { type: "@dfinity/rust@v3.3.0" } instead of writing shell commands in build.steps — the Rust recipe defaults the Cargo package to the canister name, so no configuration is needed when they match (see Pitfall 21).
Not committing .icp/data/ to version control. Mainnet canister IDs are stored in .icp/data/mappings/<environment>.ids.json. Losing this file means losing the mapping between canister names and on-chain IDs. Always commit .icp/data/ — never delete it. Add .icp/cache/ to .gitignore (it is ephemeral and rebuilt automatically). If you have environments using a connected network that gets reset frequently you can add those specific environment mapping files to .gitignore. Never add the entire .icp or .icp/data directory to gitignore.
Using icp identity use instead of icp identity default. The dfx command dfx identity use <name> became icp identity default <name> (setter). icp identity default with no argument is the getter — it prints the current default identity, equivalent to dfx identity whoami. The command icp identity use does not exist. Similarly, dfx identity get-principal became icp identity principal, and dfx identity remove became icp identity delete.
Confusing networks and environments. A network is a connection endpoint (URL). An environment combines a network + canisters + settings. You deploy to environments (-e), not networks. Multiple environments can target the same network with different settings (e.g., staging and production both on ic).

Writing networks or environments as a YAML map instead of an array. Both networks and environments are arrays of objects in icp.yaml, not maps:

# Wrong — map syntax
networks:
  local:
    mode: managed
environments:
  staging:
    network: ic

# Correct — array syntax
networks:
  - name: local
    mode: managed
environments:
  - name: staging
    network: ic
    canisters: [backend, frontend]

Forgetting that local networks are project-local. Unlike dfx which runs one shared global network, icp-cli runs a local network per project. You must run icp network start -d in your project directory before deploying locally. The local network auto-starts with system canisters and seeds accounts with ICP and cycles. Stop it when done:
```
icp network start -d  # start background network
icp deploy            # build + deploy + sync
icp network stop      # stop when done
```

Not specifying build commands for asset canisters. dfx automatically runs npm run build for asset canisters. icp-cli requires explicit build commands in the recipe configuration:

canisters:
  - name: frontend
    recipe:
      type: "@dfinity/asset-canister@v2.2.1"
      configuration:
        dir: dist
        build:
          - npm install
          - npm run build

Expecting output_env_file or .env with canister IDs. dfx writes canister IDs to a .env file (CANISTER_ID_BACKEND=...) via output_env_file. icp-cli does not generate .env files. Instead, it injects canister IDs as environment variables (PUBLIC_CANISTER_ID:<name>) directly into canisters during icp deploy. Frontends read these from the ic_env cookie set by the asset canister. Remove output_env_file from your config and any code that reads CANISTER_ID_* from .env — frontends use the ic_env cookie, and canister code reads the same variables at runtime (see Canister Environment Variables below and Pitfall 22).
Expecting dfx generate for TypeScript bindings. icp-cli does not have a dfx generate equivalent. Use @icp-sdk/bindgen (>= 0.3.0) with @icp-sdk/core (>= 5.0.0 — there is no 0.x or 1.x release) to generate TypeScript bindings from .did files at build time. Use outDir: "./src/bindings" so imports are clean (e.g., ./bindings/backend). The .did file must exist on disk — either commit it to the repo, or generate it with icp build first (recipes auto-generate it when candid is not specified). See references/binding-generation.md for the full Vite plugin setup.
Passing { agent } to createActor from @icp-sdk/bindgen. The old @dfinity/agent pattern was createActor(canisterId, { agent }). The @icp-sdk/bindgen pattern is createActor(canisterId, { agentOptions: { host, rootKey } }) — the binding creates the agent internally. Passing { agent } to the new API silently creates an anonymous identity — no error is thrown, but calls return empty data or access denied. See references/binding-generation.md for the correct pattern.
Mixing canister-level fields across config styles. When using a recipe, the only valid canister-level fields are name, recipe, sync, settings, and init_args. Fields like candid, build, or wasm are not valid at canister level alongside a recipe — recipe-specific options go inside recipe.configuration. When using bare build (no recipe), valid canister-level fields are name, build, sync, settings, and init_args. The field init_arg_file does not exist — use init_args.path instead (e.g., init_args: { path: ./args.bin, format: bin }). For the authoritative field reference, consult the icp-cli configuration reference.
```
# Wrong — candid is not a canister-level field when using a recipe
canisters:
  - name: backend
    candid: backend/backend.did
    recipe:
      type: "@dfinity/rust@v3.3.0"

# Correct — candid goes inside recipe.configuration
canisters:
  - name: backend
    recipe:
      type: "@dfinity/rust@v3.3.0"
      configuration:
        candid: backend/backend.did
```
Placing mops.toml where mops cannot find it. mops searches upward from the build working directory. Where to place mops.toml depends on how the canister is defined:
- Inline canisters (defined directly in icp.yaml): build cwd is the project root. Place mops.toml at the project root next to icp.yaml. A mops.toml in src/backend/ will not be found.
- Path-based canisters (referenced via canisters/* or ./my-canister, each with its own canister.yaml): build cwd is the canister directory. Place mops.toml in each canister’s directory for per-canister dependencies and compiler versions, or omit it to fall back to a shared mops.toml in a parent directory.
When mops.toml is not found, mops build fails because it cannot locate the project configuration. When mops.toml exists but is missing the matching [canisters.<name>] entry, see Pitfall 17.
Misunderstanding Candid file generation with recipes. Binding generation tools (e.g. @icp-sdk/bindgen) require a .did file at a known path on disk. Where to configure it depends on the recipe:

Rust — candid goes inside recipe.configuration in icp.yaml:
- If specified: the file must already exist. The recipe uses it as-is and does not generate one.
- If omitted: the recipe auto-generates the .did via candid-extractor into the build cache (no predictable project path).
To generate and commit it, then add candid: backend/backend.did inside recipe.configuration:
```
cargo install candid-extractor  # one-time setup
icp build backend
candid-extractor target/wasm32-unknown-unknown/release/backend.wasm > backend/backend.did
```
Motoko (v5 recipe) — mops build auto-generates the .did to .mops/.build/<name>.did.
- No binding generation needed — nothing to do. The generated .did in .mops/.build/ is sufficient; do not commit it.
- Binding generation needed — generate a curated .did at a stable path with mops generate candid:
```
mops generate candid backend
```
  This extracts the Candid interface directly from Motoko source without compiling WASM. With [canisters.backend].candid set in mops.toml, it overwrites that file; otherwise it writes <name>.did next to main (e.g. main = "src/backend/main.mo" → src/backend/backend.did) and records the path in mops.toml. Point the binding tool’s config (e.g. @icp-sdk/bindgen’s didFile) at that .did. Re-run after any interface change, and commit the .did and mops.toml together. Load mops-cli for details. (The older mops build backend + cp .mops/.build/backend.did … two-step still works but builds the full WASM just to extract the interface.)
Missing or mismatched [canisters] key in mops.toml. The @dfinity/motoko@v5+ recipe calls mops build <canister-name>, where the name comes from the name field in icp.yaml. mops build requires a matching [canisters.<name>] entry in mops.toml. If the entry is absent or the key does not exactly match (including casing), the build fails with:
```
No Motoko canisters found in mops.toml configuration
```
Add the matching entry — the key must equal the name: value in icp.yaml:
```
[canisters.backend]
main = "src/backend/main.mo"
```
Port 8000 already in use when starting the local network. Two scenarios:

Scenario A — another icp-cli project holds the port. Stop that project’s network using --project-root-override (a global flag available on all commands):
```
icp network stop --project-root-override /path/to/other-project
```
To run both networks at once instead of stopping one — e.g. parallel git worktrees — set gateway.port: 0 so each gets a free port. See “Parallel local networks (git worktrees)” under How It Works.

Scenario B — a non-icp service holds the port. Configure an alternate port in icp.yaml and read the actual URLs dynamically via icp network status --json rather than hardcoding localhost:8000:
```
networks:
  - name: local
    mode: managed
    gateway:
      port: 8001
```
```
icp network status --json  # returns gateway URL, replica URL, etc.
```
icp new hangs in CI without --silent. Without --define flags, icp new launches an interactive prompt that blocks indefinitely in non-interactive environments. Always pass --subfolder, --define, and --silent for scripted use:
```
icp new my-project --subfolder rust --define project_name=my-project --silent
```
Using the anonymous identity on mainnet. The local network seeds all managed identities — including the anonymous identity, which is the default — with ICP and cycles on start, so local development works out of the box with no identity or cycles setup required. On mainnet this does not apply, and the anonymous identity should never be used: it is shared by anyone, meaning ICP sent to it is publicly accessible and canisters deployed under it are uncontrolled.

Before deploying to mainnet, switch to a named identity:
```
icp identities list                                        # check available identities
icp identity default my-identity                           # switch to an existing one
# or: icp identity new my-identity && icp identity default my-identity
```
Then verify it has funds — a new identity will need to be funded with ICP or cycles before proceeding:
```
icp token balance -n ic    # check ICP balance on mainnet
icp cycles balance -n ic   # check cycles balance on mainnet
icp identity account-id    # get account ID to fund if needed
```

Over-specifying (or wrongly omitting) package in the Rust recipe. As of @dfinity/rust@v3.3.0 the package config is optional and defaults to the canister name in icp.yaml. By convention, keep the canister name equal to the [package] name in Cargo.toml and omit package entirely:

# Preferred — name matches the Cargo package, so no package config
canisters:
  - name: backend        # matches [package] name = "backend" in Cargo.toml
    recipe:
      type: "@dfinity/rust@v3.3.0"

# Required only when the names differ (e.g. a workspace crate)
canisters:
  - name: backend
    recipe:
      type: "@dfinity/rust@v3.3.0"
      configuration:
        package: my-project-backend   # actual [package] name in Cargo.toml

Omitting package when the names differ causes cargo build to fail with a package-not-found error — set it to the exact [package] name. On recipe versions before v3.3.0, package was required in all cases.

Hand-wiring canister IDs with setter methods or deploy scripts. Controller-only setters (setBridge(principal)) called by a post-deploy script — or sibling IDs hardcoded in settings.environment_variables or init args — are unnecessary: icp deploy injects every canister’s ID into every canister’s settings as PUBLIC_CANISTER_ID:<canister-name>, readable by canister code at runtime with the correct per-environment value (see Canister Environment Variables). Setter wiring is also more fragile: a --mode reinstall silently wipes the stored pointer, while the automatic variables are re-stamped on every deploy.

How It Works

Project Creation

icp new scaffolds projects from templates. Pass --subfolder, --define, and --silent for non-interactive use:

icp new my-project --subfolder rust --define project_name=my-project --silent

Available templates and options: dfinity/icp-cli-templates.

Build → Deploy → Sync

Source Code → [Build] → WASM → [Deploy] → Running Canister → [Sync] → Configured State

icp deploy runs all three phases in sequence:

Build — Compile canisters to WASM (via recipes or explicit build steps)
Deploy — Create canisters (if new), apply settings, install WASM
Sync — Post-deployment operations via script or plugin steps (e.g., uploading assets). Asset uploading is not built into the CLI: the @dfinity/asset-canister@v2.2.1 recipe supplies a plugin sync step that uploads the dir contents. The legacy built-in type: assets step is removed in icp-cli 0.3.0 — see the asset-canister skill.

Run phases separately for more control:

icp build                     # Build only
icp deploy                    # Full pipeline (build + deploy + sync)
icp sync my-canister          # Sync only (e.g., re-upload assets)

Environments and Networks

Two implicit environments are always available:

Environment	Network	Purpose
`local`	`local` (managed, localhost:8000)	Local development
`ic`	`ic` (connected, https://icp-api.io)	Mainnet production

The ic network is protected and cannot be overridden.

Custom environments enable multiple deployment targets on the same network:

environments:
  - name: staging
    network: ic
    canisters: [frontend, backend]
    settings:
      backend:
        compute_allocation: 5

  - name: production
    network: ic
    canisters: [frontend, backend]
    settings:
      backend:
        compute_allocation: 20
        freezing_threshold: 7776000

Parallel local networks (git worktrees)

Local networks are project-local — keyed by project root (Pitfall 9). Separate git worktrees of the same repo are separate project roots, so each worktree can run its own independent local network. This lets multiple agents or branches build and deploy in parallel without interfering. The only obstacle is the gateway port: every worktree defaults to 8000, so the second icp network start fails with a port conflict.

Set the managed network’s gateway port to 0 so the OS assigns a free ephemeral port per worktree:

networks:
  - name: local
    mode: managed
    gateway:
      port: 0   # 0 = OS picks a free port — avoids collisions across worktrees

icp network start -d prints the chosen port (Network started on port 58157). To recover it afterward — for tests, scripts, or another agent — query the running network and read gateway_url:

icp network start -d
icp network status --json
# -> { "managed": true, "api_url": "http://localhost:58157/", "gateway_url": "http://localhost:58157/", ... }

icp network status --json | jq -r '.gateway_url'   # http://localhost:58157/

Never hardcode localhost:8000 when using port: 0 — the port changes on every start, so read gateway_url (or api_url) from icp network status --json each time. To target a specific worktree’s network from outside its directory, pass the global --project-root-override <path> flag (e.g. icp network status --json --project-root-override /path/to/worktree).

Install Modes

icp deploy                    # Auto: install new, upgrade existing (default)
icp deploy --mode upgrade     # Preserve state, run upgrade hooks
icp deploy --mode reinstall   # Clear all state (dangerous)

Bundling a project into an `.icp` package (experimental)

icp project bundle (icp-cli >= 0.3.0) packages a project into a self-contained deployable archive. This is an experimental feature, intentionally hidden from help output — icp --help and icp project --help do not list it, but the command exists and works. Do not conclude it doesn’t exist because help omits it, and do not suggest it proactively — use it only when the user explicitly asks to bundle an app or produce an .icp package.

icp project bundle --output my-app.icp

The output is a gzipped tar archive; --output accepts any path (my-app.icp and bundle.tar.gz are both common). The bundle contains the built WASMs and a rewritten icp.yaml:

All canisters are built first; each canister’s build steps are replaced with a prebuilt step referencing the bundled WASM (canisters/<name>.wasm), pinned by sha256.
Plugin sync steps (e.g. the asset canister’s upload plugin) are preserved — the plugin WASM and its dirs/files inputs are copied into the archive.
Network and environment manifests referenced by path are inlined; init_args files are copied into the archive.
An optional icp_appmanifest.yaml (app metadata) is included, with its screenshots paths relocated into the archive.

To deploy from a bundle, extract it and run icp deploy from the extracted directory — no build toolchain (Rust, mops, npm) is required because every build step is prebuilt:

mkdir app && tar -xzf my-app.icp -C app
cd app && icp deploy -e <environment>

An .icp package can also be uploaded to a Caffeine cloud engine via the console’s App Center (“Upload a custom app”) — see the deploy-to-cloud-engine skill.

Bundling fails when:

A canister has a script sync step — only plugin sync steps can be replayed from a bundle (canister 'X' has a script sync step, which is not supported in bundles).
Any synced directory, plugin file, init_args file, or screenshot resolves outside the project directory.
The --output path is inside a directory the bundle would sync (the partial archive would include itself).
A managed network defines a bind mount with an absolute host path — bundles require relative paths for portability.

Configuration

Rust canister

canisters:
  - name: backend
    recipe:
      type: "@dfinity/rust@v3.3.0"
      configuration:
        candid: backend.did  # optional — if specified, file must exist (auto-generated when omitted)
        # package: my-crate  # optional — only when the Cargo [package] name differs from the canister name

By convention the canister name must match the [package] name in Cargo.toml. When they match, package is optional (recipe >= v3.3.0 defaults it to the canister name). Set package explicitly only when the Cargo package name differs — e.g. a workspace with a differently named crate (see Pitfall 21).

Motoko canister

The v5 recipe delegates compilation to mops build. Canister configuration (main, candid, args) moves from icp.yaml to mops.toml:

# icp.yaml
canisters:
  - name: backend
    recipe:
      type: "@dfinity/motoko@v5.0.0"

# mops.toml
[toolchain]
moc = "1.9.0"

[canisters.backend]
main = "src/backend/main.mo"
candid = "backend.did"  # optional — auto-generated to .mops/.build/ when omitted

The canister name (backend) must exactly match between icp.yaml and mops.toml. No recipe.configuration block is needed in icp.yaml.

Asset canister (frontend)

canisters:
  - name: frontend
    recipe:
      type: "@dfinity/asset-canister@v2.2.1"
      configuration:
        dir: dist
        build:
          - npm install
          - npm run build

For multi-canister projects, list all canisters in the same canisters array. icp-cli builds them in parallel. There is no dependencies field — use Canister Environment Variables for inter-canister communication.

Custom build steps (no recipe)

When not using a recipe, only name, build, sync, settings, and init_args are valid canister-level fields. There are no wasm, candid, or metadata fields — handle these in the build script instead:

WASM output: copy the final WASM to $ICP_WASM_OUTPUT_PATH
Candid metadata: use ic-wasm to embed candid:service metadata
Candid file: the .did file is referenced only in the ic-wasm command, not as a YAML field

canisters:
  - name: backend
    build:
      steps:
        - type: script
          commands:
            - cargo build --target wasm32-unknown-unknown --release
            - cp target/wasm32-unknown-unknown/release/backend.wasm "$ICP_WASM_OUTPUT_PATH"
            - ic-wasm "$ICP_WASM_OUTPUT_PATH" -o "$ICP_WASM_OUTPUT_PATH" metadata candid:service -f backend/backend.did -v public --keep-name-section

Available recipes

Recipe	Type string	Required config	Optional config
Rust	`@dfinity/rust@v3.3.0`	—	`package` (defaults to canister name), `candid`, `locked`, `shrink`, `compress`, `metadata`
Motoko	`@dfinity/motoko@v5.0.0`	—	`shrink`, `compress`, `metadata`
Asset	`@dfinity/asset-canister@v2.2.1`	`dir`	`build`, `version`
Prebuilt	`@dfinity/prebuilt@v1.0.0`	`wasm`	`sha256`, `candid`, `shrink`, `compress`

Verify latest recipe versions at dfinity/icp-cli-recipes releases. Use icp project show to see the effective configuration after recipe expansion.

Canister Environment Variables

During icp deploy, icp-cli injects the ID of every canister in the environment into every canister’s settings as PUBLIC_CANISTER_ID:<canister-name> (the principal in text form). These are canister settings env vars, readable at runtime by canister code, with per-environment values refreshed on every deploy — the same code is correct on local, staging, and mainnet with zero configuration.

Use these for inter-canister wiring instead of setter methods, init args, or post-deploy scripts (Pitfall 22). Read them lazily at call time — Motoko: Runtime.envVar<system>(...) (motoko-core v2.1.0+), Rust: ic_cdk::api::env_var_value(...) — see references/canister-env-vars.md for code and the first-install/reinstall behavior.

Frontend (JavaScript): asset canisters expose the variables through the ic_env cookie, set on all HTML responses. Read it with @icp-sdk/core:

import { safeGetCanisterEnv } from "@icp-sdk/core/agent/canister-env";

const backendId = safeGetCanisterEnv()?.["PUBLIC_CANISTER_ID:backend"];

Caveats:

Variables are only updated at deploy time. After adding a canister, run icp deploy (without a canister name) so every canister gets the complete ID set.
User-defined variables (settings.environment_variables in icp.yaml) are static YAML strings shared across environments: a create-class install (first install or --mode reinstall) re-applies them, overwriting anything set with icp canister settings update --add-environment-variable. Re-assert environment-specific values (e.g. production origins) after such a deploy.
The replica caps an environment-variable value at 128 characters.

Full reference: icp-cli environment variables.

Web identity flows

Two specialized skills cover icp identity link web. Load the right one for the task — both document the stdin “Press Enter” block, the browser sign-in step, and their flag-specific pitfalls:

deploy-to-cloud-engine — link the CLI to a cloud engine console with --auth <console-origin>, then deploy to the engine’s subnet
agent-web-identity — obtain a delegation for an app-specific principal with --app <domain>, then make canister calls as the user

Additional References

For the complete CLI and configuration schema, consult the icp-cli documentation index.

For detailed guides on specific topics, consult these reference files when needed:

references/binding-generation.md — TypeScript binding generation with @icp-sdk/bindgen (Vite plugin, CLI, actor setup)
references/canister-env-vars.md — reading PUBLIC_CANISTER_ID:<name> from canister code at runtime (Motoko/Rust examples, lazy-read pattern)
references/dev-server.md — Vite dev server configuration to simulate the ic_env cookie locally. Important: wrap getDevServerConfig() in a command === "serve" guard so it only runs during vite dev, not vite build.
references/dfx-migration.md — Complete dfx → icp migration guide (command mapping, config mapping, identity/canister ID migration, frontend package migration, post-migration verification checklist)