Clerk CLI

The clerk binary is a pre-authenticated gateway to Clerk's Backend API and Platform API, plus project-level tooling (auth, linking, env pulls, instance config). When the user asks anything that touches a Clerk resource, reach for clerk first instead of hand-rolling curl.

This skill was installed by clerk init (or clerk skill install) and is pinned to clerk {{CLI_VERSION}}. If clerk --version disagrees, refresh it with clerk skill install (or bunx clerk@{{CLI_VERSION}} skill install). The binary is always the source of truth, so run clerk <command> --help to verify anything this skill claims.

Execution environment (prefer the host, understand the sandbox warning)

Most AI coding agents default to running shell commands in a sandbox where the user's home directory, OS keychain, browser launch, localhost callback binding, or network access may be blocked. The Clerk CLI depends on all of those host capabilities:

• OS credential store: clerk auth login stores the OAuth token in the system keychain. A sandbox without keychain access reports "not logged in" even when the host is authenticated.
• Home-directory Clerk state: saved config, cached metadata, and fallback credentials live under the user's Clerk config/data directories.
• Linked project metadata: resolved from the repo's git remote plus Clerk config. Sandboxes with stripped repo state or blocked home-dir reads can misreport "not linked".
• Local .env* files: publishable and secret keys materialized by clerk env pull.
• Outbound network access to Clerk: every Backend and Platform API call.
• Browser + localhost OAuth callback: clerk auth login needs both.

In agent mode, the CLI now does a best-effort warn-once check at the host-sensitive library boundaries. When it detects that host-only Clerk state or system capabilities are unavailable, it emits:

Host-only Clerk state or system capabilities may be unavailable in agent mode. This may be a sandboxed run.
Re-run this command on the host shell before trusting auth, link, env, or API failures.

Treat that warning as authoritative. The command may continue, but any auth, link, env, config, API, browser, or OAuth callback failure from that invocation is untrusted until you rerun the same command on the host.

Prefer these commands on the user's host shell, not in a sandbox:

clerk doctor, clerk whoami, clerk auth login, clerk link, clerk env pull, clerk apps ..., clerk config ..., clerk api ....

If a command was accidentally run in a sandbox and it reports Not logged in, auth_required, not linked, missing env, keychain/file permission errors, or network failures, do not treat the result as authoritative. Rerun it on the host before acting on it or reporting it to the user.

Invoking the CLI

Before running any clerk command, figure out which binary to invoke and bind that choice for the rest of the session:

# 1. Prefer a globally installed binary when it matches the skill's pinned version.
command -v clerk >/dev/null 2>&1 && clerk --version

If that prints {{CLI_VERSION}} (or any version you trust), use bare clerk for the rest of the session.

Otherwise fall back to a package runner, in this order (matches the CLI's own preferredRunner logic, which prefers the runner that matches the project's lockfile):

Yarn Classic (v1) has no dlx; treat those projects as "no preferred runner" and fall back to the first runner from the list above that's on PATH.

The published npm package is clerk, not @clerk/cli. Never teach npm install -g clerk as the primary path. The bundled skill is versioned alongside the binary, so a globally installed mismatched version will drift. If clerk --version disagrees with {{CLI_VERSION}}, either upgrade the global install or fall back to the pinned-runner form above.

Prerequisites (run at session start)

Before running any other Clerk command in a session, verify the CLI is authenticated, linked, and healthy:

clerk --version               # confirm the binary is on PATH
clerk doctor --json           # structured health check; exit 1 if anything failed

Always run clerk doctor --json first. It catches the common setup failures (not logged in, project not linked, missing keys, stale CLI version) up front, so later commands don't fail with confusing errors. In agent mode it also includes a Host execution check that warns when Clerk's host-side config / credential directories are not writable, which is the canonical signal that the current invocation is likely sandboxed.

Each result has name, status (pass/warn/fail), message, optional detail, optional remedy (how to fix it), and optional fix (label for auto-fixable issues). Parse that and act on it, or surface it to the user. If Host execution warns, rerun the command on the host before trusting any auth/link/env/API failures from the same sandboxed run. Rerun clerk doctor --json whenever a later command starts misbehaving.

If clerk --version reports a newer CLI than this skill (which is pinned to {{CLI_VERSION}}), run clerk skill install to refresh the bundled skill. The CLI binary is always the source of truth.

The mental model

A project is "linked" to an application via clerk link. Once linked, most commands auto-resolve the target app and dev instance from the repo's git remote. To target something else, pass --app <id> and/or --instance dev|prod|<instance_id>. See references/auth.md for the full resolution order.

Discover endpoints — don't memorize them

The CLI ships with the Clerk OpenAPI catalog. Always discover endpoints dynamically instead of guessing paths:

clerk api ls                  # list every Backend API endpoint
clerk api ls users            # filter by keyword (matches path, summary, tag, operationId)
clerk api ls --platform apps  # list Platform API endpoints

Use this before clerk api <path>. If you don't see the endpoint you expected, it probably isn't exposed.

The clerk api command (the workhorse)

clerk api makes authenticated HTTP calls. It auto-resolves keys, auto-detects method from body presence, supports stdin, and can preview mutations with --dry-run.

# GET requests
clerk api /users                                  # list users
clerk api /users/user_abc123                      # fetch one
clerk api /users?limit=5&order_by=-created_at     # query params work inline

# Mutating requests
clerk api /users -d '{"email_address":["[email protected]"]}'          # POST (auto-detected from body)
clerk api /users/user_abc123 -X PATCH -d '{"first_name":"A"}'
clerk api /users/user_abc123 -X DELETE

# Body from file or stdin
clerk api /users --file payload.json
cat payload.json | clerk api /users

# Always preview mutations first
clerk api /users/user_abc123 -X DELETE --dry-run
clerk api /users/user_abc123 -X DELETE --yes      # skip confirmation once you've verified

# Target a specific app/instance
clerk api /users --app app_abc123 --instance prod

# Include response headers when debugging
clerk api /users --include

# Platform API (account-level, not tenant data)
clerk api /v1/platform/applications --platform

For instance config, prefer the dedicated clerk config ... commands over raw Platform API /config paths. They handle dry-run, diffing, and confirmation more cleanly than the raw endpoint form.

Always --dry-run a mutation before running it for real. Then re-run without --dry-run (add --yes if you're sure). In agent mode, interactive confirmation is bypassed, so --dry-run is the only safety net for destructive calls.

JSON bodies must be valid JSON. The CLI validates and rejects malformed payloads.

Endpoint paths may be given with or without /v1/ prefix — both work for Backend API calls. The CLI normalizes.

See references/recipes.md for concrete patterns: listing/filtering users, creating orgs, impersonation sessions, etc.

Inspecting large outputs (do not flood your context)

users list, apps list, config pull, and most clerk api GETs return payloads that can be many kilobytes or megabytes. Production tenants commonly have thousands of users; an instance config can be hundreds of fields deep. Reading those responses into the conversation costs context window for no benefit. Save the response to a file first, then query just what you need with jq:

# 1. Persist the response. Use --limit 250 to maximize page size for users list.
clerk users list --json --limit 250 > /tmp/users.json
clerk apps list --json                > /tmp/apps.json
clerk api /users/user_abc123          > /tmp/user.json

# 2. Inspect only what you need.
jq '.data | length'                       /tmp/users.json   # current page size
jq '.hasMore'                             /tmp/users.json   # are more pages available?
jq '.data[0] | keys'                      /tmp/users.json   # discover the user shape once
jq '.data[] | {id, email_addresses}'      /tmp/users.json   # project to a few fields
jq '[.data[] | select(.banned)] | length' /tmp/users.json   # aggregate without reading rows

If jq is not available, fall back to Python or Node — both can stream the file without printing it whole:

python3 -c 'import json; d=json.load(open("/tmp/users.json")); print(len(d["data"]), d["hasMore"])'
node -e 'const d=require("/tmp/users.json"); console.log(d.data.length, d.hasMore)'

cat / head the file only when you genuinely need to see the raw structure for one-off debugging. When walking pages, write each page to its own file (e.g. page-${offset}.json) so individual pages stay independently inspectable.

Core commands at a glance

clerk <command> --help is the source of truth for flags. This table is a hint, not a spec. Before running an unfamiliar command or flag combination, run clerk <command> --help once per session. Every command also defines setExamples([...]) in source, which --help renders as a copy-pasteable Examples block, so you rarely need to guess syntax.

Agent-mode behavior (important)

The CLI auto-detects agent mode when stdout is not a TTY, or when --mode agent / CLERK_MODE=agent is set. In agent mode:

• Interactive prompts are disabled. Commands that would normally show pickers (link without --app, unlink without --yes, users without a subcommand) either auto-resolve or exit with a usage error. clerk api with no args prints usage guidance and exits 0; pass an endpoint (or ls) explicitly. Always pass explicit flags (--app, --yes) in scripted calls.
• Host-sensitive operations emit a sandbox warning once per invocation. Home-directory Clerk state, keychain access, networked Clerk calls, browser launch, and localhost OAuth callback setup can trigger the warning shown above. If it appears, rerun the same command on the host before trusting the result.
• If your harness does not clearly present as agent mode, force it. Use --mode agent or CLERK_MODE=agent when you want the CLI's non-interactive behavior and sandbox warning path to apply deterministically.
• link supports deterministic agent flows. In agent mode, clerk link --app <id> links directly. Without --app, the CLI will try silent key-based autolink first; if it cannot determine the app unambiguously, it exits and tells you to pass --app.
• init never selects or creates a real Clerk app for you in agent mode unless authenticated or given a target. Pass --app <id> (or pre-link the project) to authenticate and link a real app, or pass --keyless to use auto-generated temporary development keys when bootstrapping a new project on a keyless-capable framework. Without either, agent mode prints manual setup guidance and exits cleanly.
• unlink requires --yes in agent mode. This preserves the same safety bar as other destructive commands while still letting an agent complete the unlink non-interactively.
• Mutations still require --yes unless you accept per-call confirmation is impossible.
• doctor --fix is ignored. Parse doctor --json output's remedy field and act on it yourself.
• apps list and apps create default to JSON when piped.
• users defaults to JSON when piped, like apps. clerk users list and clerk users create emit JSON in agent mode. Bare clerk users (no subcommand) is a usage error in agent mode — pass list, create, or open explicitly. clerk users open requires the user-id positional in agent mode and prints a JSON descriptor instead of launching a browser.
• --input-json <json|@file|-> expands JSON into flags on any command (e.g. clerk init --input-json '{"framework":"next","yes":true}'). Piped stdin is also accepted: echo '{"yes":true}' | clerk init. Place --input-json after the leaf subcommand. Full rules in references/agent-mode.md.

Full matrix and sandbox details in references/agent-mode.md.

Output format and errors

• JSON output: --json on apps list and doctor. For clerk api, the response body is the raw API JSON, so pipe into jq freely.
• Exit codes: 0 success, 1 runtime error, 2 usage/validation error. doctor returns 1 if any check failed.
• Error format: User-facing errors print a single line to stderr and set a non-zero exit code. Use --verbose for stack traces when debugging.

Safety rules for autonomous use

1. Discover before acting: clerk api ls <keyword> before clerk api <path>.
2. Preview mutations: --dry-run on every config patch, config put, api -X POST/PATCH/PUT/DELETE.
3. Target explicitly in production: pass --instance prod rather than relying on defaults, and confirm with the user before any production mutation.
4. Never commit secrets: env pull writes to .env.local (which should be gitignored). Don't paste secret keys into code or chat.
5. Use doctor --json to diagnose before assuming the CLI is broken.

References

• references/auth.md — auth flow, key resolution order, host-vs-sandbox behavior, --app/--instance targeting, Backend vs Platform API.
• references/recipes.md — copy-pasteable recipes for common Clerk tasks.
• references/agent-mode.md — agent-mode behavior matrix, sandbox warning semantics, exit codes, error format.