/jfrog-init — verify and guide JFrog plugin readiness
First output must be a tool call, not text. No "I'll start..." preamble.
Walks a fixed, ordered checklist and stops at the first red result, guiding
the user through the matching fix before re-checking. Every detector in
scripts/ is idempotent, read-only, JSON-emitting, and implemented in
Node (.mjs), so the same detector runs unmodified on macOS, Linux, and
Windows. Step 3's web-login uses this skill's own local .mjs scripts,
same as every other step. The one interactive fix step that isn't a Node
script at all is Step 1's Node install itself — a missing Node can't run
a .mjs installer, so that one step branches on OS or shell directly.
Approval model — every user-approval point is detailed in its own Step
below (exact wording, AskUserQuestion payloads, forbidden phrases);
this is just the map:
node — AskUserQuestion Yes/No before auto-installing (Step 1).
jf (JFrog CLI) — AskUserQuestion Yes/No before auto-installing
(Step 2).
jf config — AskUserQuestion picker (web login, in-session, vs.
access token, which never enters this conversation) (Step 3/4).
Project selection — the user answers with a project name or key;
an AskUserQuestion picker offers the first two enumerated projects
plus "Other" (Step 6).
Everything else in the walk is read-only, except Step 5's placeholder
substitution (writes the plugin's mcp.json, unattended by design —
no AskUserQuestion, see references/script-invocation.md), Step 8's
~/.netrc write (also unattended, no AskUserQuestion — see
references/marketplace-setup.md), and the Final summary's state
write below.
Step 1's nvm install and jfrog-install-jf-cli.mjs, Step 3's
web-login scripts, Step 8's jfrog-add-claude-marketplace.mjs call,
and the Final summary's jfrog-state-file.mjs set call are
deliberately not in allowed-tools and will raise the harness's
own approval prompt — intended, not a misconfiguration; see
references/script-invocation.md.
${CLAUDE_SKILL_DIR} below is this file's own directory. Claude Code
substitutes it automatically, identically, in both this text and the
Bash rules above — write it literally rather than
resolving it yourself, so the two stay byte-for-byte consistent
regardless of install depth (see ). On
a harness that doesn't perform this substitution (e.g. Cursor or Codex,
neither of which consults for approval — every command
below still raises its own prompt there), replace it with the real
absolute path of this file's directory yourself, same as before.
allowed-tools
references/script-invocation.md
allowed-tools
At a glance (always-read core)
Order matters. Walk Steps 1–8
in exact order, stop at the first non-green result — except Step 5
red/error, Step 6's one-retry cap, Step 7's "not entitled" and
"catalog unreachable" outcomes, and Step 8 entirely (all four
non-blocking). See
The checklist, in order.
rc=$? is mandatory on every detector invocation — see
Invoking scripts. A
bare ; true hides every red/ask result as green.
Approval gates:AskUserQuestion Yes/No before auto-installing
Node (Step 1) or jf (Step 2); AskUserQuestion picker for
web-login vs. token (Step 3/4); AskUserQuestion picker for project
selection (Step 6). Everything else is read-only except Step 5's
placeholder substitution (plus, for kiro-cli, creating or merging the
jfrog entry into ~/.kiro/settings/mcp.json), Step 8's ~/.netrc
write, and the Final summary's state write.
Never surface the checklist. Run silently — no step narration, no
raw JSON/exit codes, no branch-reasoning said out loud. See
Customer-facing output.
<server-id> for Steps 4-7 always comes from the shared resolver
(explicit arg → JF_SERVER_ID → isDefault → sole configured server
→ ask) — never invented, never jf's own fallback. Step 8 reuses the
same value. See
Resolving <server-id>.
Persist state before the final summary — run
jfrog-state-file.mjs set whenever Steps 1-4 are green, regardless of
Step 5/6/7. See Final summary.
Never store, log, or print an access token — credentials stay
inside jf's own process or in-memory for one fetch call. Step 8
is the one deliberate exception (writes ~/.netrc) — see
Step 8 and
Non-goals.
This skill is the exception to, not a consumer of, the base
../jfrog/SKILL.md's prerequisites — do not
run its environment check as a gate before starting this walk.
Read the base ../jfrog/SKILL.md for foundational
context — JFrog Platform concepts and terminology, and the
Server selection rules
this skill's own "Resolving <server-id>" section (below) follows the
same never-guess/never-infer philosophy of, via its own mechanism.
This skill is the exception that makes the base skill's prerequisites
true, not a consumer of them. The base skill's
environment check assumes jf is
already installed at a working version — /jfrog-init is what gets a
user from nothing installed to that point. Do not run the base skill's
environment check as a gate before starting this walk; Steps 1-4 below
are this skill's own, more granular equivalent (Node, jf CLI install,
connection, credentials) purpose-built for the "nothing works yet" case.
Deliberately does not export JFROG_CLI_USER_AGENT, unlike the
base skill's invariant — see runJf() in scripts/lib/jf.mjs for why
(telemetry-only impact).
Customer-facing output
The user does not need to see the checklist you are walking, but does
need to see what actually happened. Run the detectors silently,
capture their output for your own reasoning, and surface only what the
user needs to know or act on:
Do not narrate step numbers ("Step 1…", "moving to Step 3…")
while the walk is in progress.
Do not paste detector JSON, exit codes, or shell command output
into the reply.
Do not narrate the branch-selection reasoning behind an
AskUserQuestion or plain-text prompt — e.g. explaining that
unresolved wasn't "server", or that candidatesWithNames had two
or more entries, so this is "the generic ask using the first two
candidates." That reasoning (in server-picker.md, project-picker.md,
and the other reference docs' branch tables) is written for you to
follow silently, not to summarize out loud — the field names in it are
never user-facing. The only output the user sees at an ask point is
the prompt itself.
Do not narrate whether the AskUserQuestion tool is available in
the current harness before falling back to the plain-text prompt
(e.g. "the AskUserQuestion tool isn't available here, I'll present
this as a plain question instead"). If it isn't available, silently
use the plain-text fallback already documented for that ask point.
Do not announce that you're about to run the checklist, or name
which check comes first — not even generically ("I'll run the setup
checklist silently, starting with the JFrog CLI check" is itself a
violation: it names a step while claiming to be silent). The same
applies to reading reference docs: "I'll start by reading the flow
docs" is a preamble. Silently means no preamble message at all — not
before running commands, not before reading files. Say nothing until
you have something the user needs to act on (an ask, a red result)
or the final summary.
Instead:
When everything passes, give a short recap in the final summary
(see "Final summary" below): a short, emoji-based checklist — JF CLI
& Config, JFrog MCP Plugin, Project & AI Catalog — so the user sees
the end state of every check at a glance, not raw step numbers and
not the Node.js check (an implementation detail, not user-facing).
When something is red, say what's wrong in plain English and
what the user needs to do next, in one or two sentences. Show the
exact command they need to run (they must see what they're
approving).
On failure, the raw detector error line is fair game to include
verbatim as a debugging aid — one line, without the JSON wrapper.
The rest of this file documents the flow for you (the model), not
for the user.
Invoking scripts: avoid the red "Error" framing
Every detector command shown below signals red/ask states via a
non-zero exit code, by design — append ; rc=$?; true when invoking
any of them. rc=$? is not optional: every Step's branch table
below keys off the exit code, and a bare ; true throws it away, so
every red and ask silently reads as green. Read
references/script-invocation.md in full before running any command
in this walk — the exact pattern and why it's required, not optional
background.
Flow
Follow this flow literally. Every decision node is covered by a
detector or fix script below; every user-facing prompt uses the exact
wording documented in the corresponding step. Do not reorder, do not
skip, do not narrate the diagram to the user. Read
references/flow-diagram.md for the full flowchart before starting a
walk — the same logic as the Steps below, drawn as a map.
The checklist, in order
Node.js ≥ 18 installed? — no script; run node --version / npx --version directly
AI Catalog reachable & user entitled? — scripts/jfrog-detect-catalog-runtime.mjs [server-id]
Claude agent-plugin marketplace registered? — scripts/jfrog-add-claude-marketplace.mjs [server-id] [project-key], Claude Code only
Run detectors in this exact order and stop at the first non-green
result — except Step 5 going red/error (see Step 5), Step 6 hitting
its one-retry cap (see Step 6), Step 7's "not entitled" and "catalog
unreachable" outcomes (see Step 7), and Step 8 entirely (see Step 8),
all four non-blocking. Step
1 has no script — a Node script can't verify Node exists — so every
step after it is written in Node and can assume Node is present. The
JPD URL is read directly from
jf config; there is no separate
JFROG_PLATFORM_URL env var. The project key is asked every walk in
Step 6 (the state file at ~/.jfrog/setup.json may supply a
"reuse the current project?" hint); Step 7's catalog probe takes no
project argument — it only checks catalog reachability and entitlement
for the resolved server.
Resolving <server-id> for Steps 4-7
Order (used by every detector that takes [server-id] — resolved
through the single shared scripts/jfrog-resolve-jf-server.mjs):
Explicit argument passed to the detector.
JF_SERVER_ID env var.
The server flagged "isDefault": true in ~/.jfrog/jfrog-cli.conf.v6
— resolved automatically via scripts/jfrog-resolve-jf-server.mjs.
If only one server is configured, it is used silently.
Otherwise the detector exits 2 ("ask") with a JSON candidates
list of the configured server IDs. Stop and read
references/server-picker.md in full — it has the exact
AskUserQuestion payload for showing the user the actual servers to
pick from; do not paraphrase or invent your own prompt. Never invent
a server, never rely on jf's own fallback.
Step 1: Node.js ≥ 18 installed?
node --version; true
npx --version; true
No script — Node's own binary is the only thing that can answer
"is Node installed", so there's nothing a script would check that
these two commands don't already answer directly. Checked first
because every other script in this walk — including every other
detector — is a Node program. npx matters because the JFrog MCP
server (mcpServers.jfrog, Step 5) is launched via npx @jfrog/agent-guard — no npx means the MCP entry can't start,
regardless of everything else.
Read the output yourself, no JSON to parse:
Either command errors (e.g. command not found: node) → red:
Node.js (or npx) is not installed / the install is broken.
node --version prints a version like v16.2.0 → parse the major
number yourself. < 18 → red: "Node.js <version> is too old —
jfrog-init requires Node ≥ 18."
node --version ≥ 18 andnpx --version succeeds → green →
proceed to Step 2.
Never paste the raw shell output. Translate to plain English —
"npx is not installed" not `command not found`, "Node.js v16 is
too old" not the version string verbatim. The raw output is for your
reasoning, not for the user.
On red, stop and read references/node-install-prompt.md in full
before responding to the user. It has the exact AskUserQuestion
payload, the forbidden phrases, and the install commands — required
behavior, not optional background. Even on the install path there's no
detector script: a missing Node can't run a .mjs installer, so the
install is a bash/PowerShell command the model runs directly.
Exit 1 (red), reason: "missing" → jf not found on PATH → stop
and read references/jf-cli-install-prompt.md in full — required
behavior, not optional background.
Exit 1 (red), reason: "broken" → jf is on PATH but hung or
failed to run → stop and read references/jf-cli-install-prompt.md
in full — it has a separate payload for this case; required
behavior, not optional background.
Exit 1 (red), reason: "outdated" → jf installed but below the
required minimum version → stop and read
references/jf-cli-update-prompt.md in full — required behavior,
not optional background.
Read references/jf-cli-install-internals.md for the full plan
A/B/C breakdown — why this walk uses npm install -g jfrog-cli-v2-jf
instead of the base skill's OS-specific install docs, and how the
fallback to a checksum-verified direct binary download works.
Checked here (right after installing jf) so the JFrog CLI is
finished — install + auth — before the walk moves on to any other
prereqs. Keeps the "get jf working" story in one contiguous block
from the user's point of view.
This step does not pick or validate any specific server — it only
checks that jf config show returns at least one configured server,
regardless of how many there are or whether one is marked isDefault.
If multiple servers are configured with none marked default, that
ambiguity is surfaced starting at Step 4, not here — see "Resolving
<server-id> for Steps 4-7" above.
Exit 0 (green) → proceed to Step 4.
Exit 1 (red) → jf is installed but not connected to any
server. Stop and read references/jf-config-auth-picker.md in
full before responding to the user — the Step 3 section has the
exact web-login-vs-token AskUserQuestion, this skill's own local
web-login script copies (no cross-skill call into the base jfrog
skill), and the token command template — required behavior, not
optional background.
Stop and read references/jf-config-auth-picker.md's "Step 4's full
branch table" section in full — the two sub-checks and the required
exit-code branches (credentials-only fix on Exit 1, Exit 2 still
blocks) — required behavior, not optional background.
Step 5: JFrog MCP plugin file has a jfrog entry?
If you are Kiro or Kiro CLI, prefix the command with
JFROG_INIT_HARNESS=kiro or JFROG_INIT_HARNESS=kiro-cli on the
same line — no other harness needs this. Silent, like every other
setup command in this walk.
Pass the same <server-id> already resolved for Step 4 (empty string
if Step 4 resolved silently via default/single-server) — this reuses
it for the placeholder fix instead of re-resolving from scratch.
Read-only against the JFrog plugin's own mcp.json — with one
exception: automatic placeholder substitution of an unresolved
${JFROG_PLATFORM_URL} / ${JFROG_URL} with the real JPD URL from
jf config (plus, for kiro-cli, adding the jfrog entry to
~/.kiro/settings/mcp.json when no plugin has put one there — see
references/mcp-plugin-config.md).
Stop and read references/mcp-plugin-config.md in full — exactly
how the substitution works, the per-harness plugin-config paths, and
the required exit-code branches (Exit 1/3 non-blocking, Exit 2 the one
outcome that still blocks) — required behavior, not optional
background.
State reuse across walks. Before asking the user for a project,
stop and read references/project-state-reuse.md in full — it has
the exact "reuse <KEY>?" AskUserQuestion and the jpdUrl-drift check
this step requires, not optional background.
Name-or-key input is resolved via exact match, then progressively
fuzzier tiers — see references/project-matching.md for the algorithm.
Picking a project, interactively. Whenever the detector needs the
user to choose — no input was passed, the typed input didn't match
anything (404), or it matched more than one project (ambiguous) —
stop and read references/project-picker.md in full before
responding to the user. It has the exact AskUserQuestion payload
shapes for each case (a 404 with close suggestions, an ambiguous or
missing input, and the no-AskUserQuestion plain-text fallback) and
the forbidden-phrasing rules for each — this is required behavior for
the step, not optional background.
Stop and read references/project-resolution-branches.md in full
for exactly how to branch on the detector's exit code (green / ask /
red with the one-retry cap / error) — required behavior, not optional
background.
Stop and read references/catalog-runtime-branches.md in full —
the two sub-checks (anonymous reachability, authenticated entitlement)
and the required exit-code branches (Exit 1 "not hosted/unreachable"
and Exit 4 "not entitled" are both non-blocking; Exit 2 still blocks) —
required behavior, not optional background.
Step 8: Claude agent-plugin marketplace registered?
Two preconditions, in this order. Step 7 must have been green — the
marketplace is served by the same AI Catalog that Step 7 probes, so
after a non-blocking red there (unreachable, or not entitled) there is
nothing to register. Then, Claude Code only — check the current
harness by reusing detectHarness() from
scripts/jfrog-resolve-mcp-config.mjs (the same export Step 5 already
uses), e.g.
node -e "import('${CLAUDE_SKILL_DIR}/scripts/jfrog-resolve-mcp-config.mjs').then(function(m){console.log(m.detectHarness())})".
If either precondition fails, skip this step silently — never run
the script below, no AskUserQuestion, no note anywhere, not even in
the Final Summary. Treat it exactly as if Step 8 didn't exist for this
walk.
Pass the same <server-id> already resolved for Step 4 (empty string
if Step 4 resolved silently via default/single-server), then Step 6's
canonical project key, or an empty string if Step 6 resolved none.
Stop and read references/marketplace-setup.md in full before
acting on the exit code — required behavior, not optional
background.
Exit 0 (green) → success. The last stdout line is
Successfully added marketplace: <marketplace-name> — extract
<marketplace-name> for the Final Summary's trailing line.
Exit 1 or 3 (red) → non-blocking failure. Nothing beyond the Final
Summary's ⚠️ line, and never volunteer which cause it was.
Final summary
Persist the walk's state before rendering any outcome below.
Whenever Steps 1-4 are green (regardless of what Step 5/6/7 reported),
run:
using the server-id resolved earlier in this walk, Step 4's own
jpdUrl field (present on its green JSON result), and Step 6's project
key — its resolvedKey on green, or "" if Step 6 never resolved one
(ambiguous input, 404, 403, or the one-retry cap was hit). This is the
only thing that writes ~/.jfrog/setup.json when the walk is followed
step-by-step; it's the same file Step 6's "reuse <KEY>?" prompt
(project-state-reuse.md) reads on a future walk, so skipping this call
means that prompt has nothing to offer next time. Skip it only if Steps
1-4 themselves didn't all pass — there's nothing resolved yet to
persist. (Running the whole walk via jfrog-detect-all.mjs instead —
see "Running everything at once" below — does this same write itself;
don't call both.)
Give the user a short recap, not the raw checklist. See
"Customer-facing output" above — no step numbers, no raw JSON. Render a
short, emoji-based checklist, not a prose paragraph or a five-line
plain-text list. Three grouped lines cover all five checks:
JF CLI & Config — Steps 2-4 (jf installed and connected to a
server). Always fully resolved here — this checklist only renders
once Steps 1-4 all passed (see "Anything else red" below for the
alternative).
JFrog MCP Plugin — Step 5.
Project & AI Catalog — Steps 6 and 7 together.
Skip Node.js (Step 1) — implementation detail, not user-facing.
Rules for the checklist:
Do not use the word "done" anywhere in it.
Keep those checks in exactly these three grouped lines — never
expand back out to five.
All three groups fully resolved → use this exact format, verbatim:
For the merged Project & AI Catalog line, if only one of the two
is outstanding name just that one; if both are, separate them with a
semicolon: ⚠️ Project & AI Catalog — project not set up yet; catalog access not entitled.
Step 8 gets a fourth checklist line, but only when it ran —
nothing appears when it was skipped:
Success — ✅ JFrog Marketplace, plus this trailing sentence
after the checklist block, in this exact wording:
Added the JFrog marketplace <marketplace-name> to Claude Code.
Browse available plugins with /plugins, or install directly with
claude plugin install <plugin>@<marketplace-name>
Red — ⚠️ JFrog Marketplace — not registered, and no trailing
sentence.
Never phrase a ⚠️ line as a failure or as something the user needs to
fix before continuing — all of them are non-blocking by design. The
short fact after the em dash is the same underlying cause this skill
has always surfaced, just worded without "pending":
Step 5 red/error (MCP plugin not configured):not configured.
If the user asks why or how to fix it, that's when the specific cause
from Step 5's detail comes in — either run
jfrog-reinstall-jfrog-plugin.mjs (see Step 5) for the per-harness
reinstall remedy, or point at resolving jf config, matching
whichever cause Step 5 actually reported.
Step 6 hit its retry cap (no project resolved):project not set up yet. If the user asks, mention they can pick one whenever they're
ready. The server/JPD are still recorded to the state file either way
(see the persistence step at the top of this section); a project
picked in an earlier walk, if any, is left as-is rather than cleared.
Step 7 returned exit 4 (not entitled):catalog access not entitled. If the user asks for the fix: ask your JFrog admin for the
"AI Catalog Read" role to browse or install MCPs from the catalog.
Step 7 returned exit 1 (catalog not hosted / unreachable):catalog not reachable on this JPD. No fix instruction; there may be
nothing to fix (this JPD may simply not host the AI Catalog).
Something happened this walk (Node.js/jf CLI installed, jf config connected, a project resolved in Step 6, an MCP placeholder
substituted in Step 5, etc.): still the same checklist — the action
itself isn't called out per-line, ✅ is ✅ regardless of whether it
needed fixing this walk.
Anything else red (Steps 1-4 not all green): one or two sentences
naming what's blocking and what to do next, no checklist — there's
nothing to check off yet. Include the raw detector error line if it
helps debug, without the JSON wrapper.
Running everything at once
Read references/batch-walk.md in full for jfrog-detect-all.mjs's
exact semantics — the non-blocking exceptions, the JSON summary
fields, and the state-file write behavior.
Non-goals (out of scope for this skill)
Installing the JFrog IDE plugin, or replacing its auto-config.
Installing the VS Code hook.
A first-MCP wizard for an empty catalog.
Persisting the picked project key to JF_PROJECT or any shell
profile. Step 6 asks every walk and threads the pick forward as a
positional argument only — nothing about project selection ever
touches a shell profile. (Two other, unrelated things in this walk
do: Step 1's nvm-based Node install, and Step 2's Plan C fallback
when npm itself isn't usable — both append one PATH line to the
user's shell rc file, disclosed up front in the install consent
prompts, see node-install-prompt.md / jf-cli-install-prompt.md.
Plans A/B of Step 2 — the common case — don't touch a shell profile
at all, relying on npm's own global bin directory instead.)
Granting AI Catalog roles/permissions — Step 7 only instructs.
Storing access tokens to disk, logging them, or printing them.
Step 4's authenticated check keeps the credential inside jf's own
process (jf rt ping); Steps 6 and 7 extract it from jf config export only in memory, for one fetch call. Step 3/4's token-based
jf config path (see references/jf-config-auth-picker.md) never
touches this skill or the model at all — the user runs that command
themselves. Step 8 is the one deliberate exception — it writes
the token to ~/.netrc; see references/marketplace-setup.md.
Walk Steps 1-8 in exact order; stop at the first non-green result
(Step 5 red/error, Step 6's retry cap, Step 7 "not entitled" or
"unreachable", and Step 8 entirely are non-blocking)
Every detector invocation appends ; rc=$?; true — never a bare
; true
AskUserQuestion before auto-installing Node (Step 1) or jf
(Step 2); picker for web-login vs. token (Step 3/4); picker for
project selection (Step 6)
Silent walk — no step narration, no raw JSON/exit codes, no
branch-reasoning surfaced to the user
<server-id> for Steps 4-8 comes only from the shared resolver —
never invented, never jf's own fallback
Steps 1-4 green → jfrog-state-file.mjs set <server-id> <jpdUrl> <project-key> before rendering the final summary
Never store, log, or print an access token — except Step 8's
~/.netrc write
Read the base ../jfrog/SKILL.md for context; do not run its
environment check as a gate before this walk
