Generates a secure, branded, organization-specific launcher that wraps Claude Code, Codex CLI, Gemini CLI, Cursor, or Cline onto a corporate AI gateway, then helps the user distribute it to their team. Trigger phrases include "corporate launcher", "wrap claude code", "wrap codex", "wrap gemini", "white-label cursor", "white-label cline", "internal AI CLI", "bedrock gateway", "azure openai cli", "vertex cli", "ship to my team", "internal copilot". Use this skill whenever a user needs to wrap a vendor CLI onto a corporate AI gateway.
Instrucciones de origen · Vista previa de solo lectura
name
corporate-launcher
description
Generates a secure, branded, organization-specific launcher that wraps Claude Code, Codex CLI, Gemini CLI, Cursor, or Cline onto a corporate AI gateway, then helps the user distribute it to their team. Trigger phrases include "corporate launcher", "wrap claude code", "wrap codex", "wrap gemini", "white-label cursor", "white-label cline", "internal AI CLI", "bedrock gateway", "azure openai cli", "vertex cli", "ship to my team", "internal copilot". Use this skill whenever a user needs to wrap a vendor CLI onto a corporate AI gateway.
when_to_use
Invoke when an employee of a regulated org (bank, telco, public sector, defense, healthcare) needs an AI coding CLI but cannot use the vendor's public endpoint. Use when a user says their company "does not authorize" Claude / Codex / Gemini / Cursor, asks for a white-label CLI for their team, needs a wrapper for AWS Bedrock / Azure OpenAI / Vertex AI / LiteLLM, must enforce corporate proxy + SSL inspection + custom CA, or wants to bundle a set of internal skills for colleagues. The deliverable is always a runnable, branded launcher plus a distribution kit — not just shell scripts or a one-off API call. Pair with a corporate gateway (Bedrock, Azure OpenAI, Vertex, LiteLLM).
telemetry off, only first-party models from the chosen provider
a turnkey install for their team (one script, one config, one launcher)
a curated set of skills (design pack, internal review skill, MCP servers) shipped inside the launcher
Anti-trigger: do not trigger for a personal/hobby setup, a one-off API call, or generic shell scripting.
Workflow (always follow in order)
Posture: This skill writes to disk and may invoke gh repo create --push. Set user-invocable: true to require explicit /corporate-launcher invocation in environments where automatic side-effects must be avoided.
Phase 0 — Confirm intent (skip if obvious)
If the user only said "wrap claude for my company" with no detail, ask 1-2 quick clarifying questions via AskUserQuestion:
Which CLI to wrap (Claude Code, Codex CLI, Gemini CLI, Aider, opencode)?
Which backend gateway (LiteLLM, Bedrock, Azure OpenAI, Vertex, custom OpenAI-compatible)?
If they already named both, skip to Phase 1.
Phase 1 — Run the config interview
Read references/interview-flow.md and walk through every section. Use AskUserQuestion for each step. Save answers to a session JSON in memory.
Branding — system prompt overrides, banner, terminal title, color theme
Skills bundle — which skills + MCP servers to include for colleagues
Distribution — how to ship the launcher to the team (public/private repo, tarball, one-liner)
If the user replies "I don't know" to a network/cyber question, mark it unknown and continue — generate sane defaults and flag them in the post-install summary so the security office can review.
Phase 1.4 — Legal compliance check (NEW v0.9)
Why: each wrapped CLI has Terms of Service that restrict which model providers it can legally talk to. The textbook example is Claude Code → OpenAI / Azure OpenAI / Gemini = breach of Anthropic Commercial Terms §D.4 ("no competing services"). Anthropic publicly revoked OpenAI's Claude API access in August 2025 citing exactly this clause. A corporate launcher that generates a breach-of-contract configuration is a Day-1 lawsuit risk for the customer org.
What: reads ${CLAUDE_SKILL_DIR}/scripts/legal-matrix.json (the result of reading each CLI's TOS, dated and re-verifiable). For each WRAPPED_CLIS × backend pair, looks up the verdict:
allowed → silent pass.
ambiguous → block unless the operator supplies --legal-reviewed=YYYY-MM-DD --legal-reviewer="Name <email>". Documents the legal review in <install>/.legal-attestation.json.
forbidden → refuse generation with the verbatim TOS citation. Override only via --legal-override="<documented reason>" (rare — requires legal counsel sign-off).
Freshness: refuses generation if legal-matrix.json is older than reverify_after_days (default 180). A PreToolUse hook also surfaces stale-matrix warnings before the user invests time in the interview.
Re-verification: re-run the TOS-reading agent every 6 months. The matrix carries last_read_date so audits know exactly when each verdict was sourced.
Phase 1.5 — Probe the gateway
Why: catches typos in URL, expired tokens, missing models, TLS issues before the user invests time in the rest of the interview.
What: invokes ${CLAUDE_SKILL_DIR}/scripts/api-probe.py with the user's URL + token. Reports reachable / unreachable, models catalog, latency, TLS cert.
If the probe fails, the skill should pause and ask: "Continue anyway, or revise the answer?" — never silently proceed with a known-broken gateway.
Phase 1.6 — Probe with load test (optional)
Gated by LOAD_TEST_ENABLED=yes. Skip otherwise.
What: invokes ${CLAUDE_SKILL_DIR}/scripts/load-test.py to run a sustained-throughput and concurrency probe against the gateway. Reports requests/sec, p50/p95/p99 latency, error rate, and concurrent-stream ceiling.
Use this when the launcher will be shipped to a large team and the RSSI/CISO needs evidence the gateway can absorb the load before sign-off.
Phase 2 — Validate
Before generating anything:
check every required field has a value (or sane default)
show a one-screen recap
ask: "Generate the launcher? [y/N]"
Don't generate without explicit confirmation.
Phase 2.5 — Inject corporate dev rules (NEW v0.6)
Why: the launcher must carry the org's coding standards (review checklists, naming conventions, security patterns) alongside the cyber baseline, so colleagues inherit them on day one.
What: invokes ${CLAUDE_SKILL_DIR}/scripts/dev-rules-installer.py to fetch the org's dev rules and write dev-rules.md next to BRANDING.md in the rendered tree. The installer supports 4 source modes:
none — skip; no dev rules file written.
inline — user pastes the rules during the interview; written verbatim.
local — read from a local path on the user's machine.
git — clone/fetch from an internal git URL (HTTPS + token or SSH).
The launcher appends dev-rules.md to its --append-system-prompt chain so every prompt carries the corporate dev conventions.
Phase 3 — Generate the launcher
For each selected CLI:
Pick the matching folder under templates/<cli-id>/
Substitute every ${VAR} with the answers via ${CLAUDE_SKILL_DIR}/scripts/render.py
Write the rendered files into the install path
Render the shared/ modules once (VPN check, proxy detect, cyber rules, cost tracker, prompt filter, strip-proxy if needed)
Render the skills bundle — clone or copy the chosen skills under <install>/skills/
Set executable bits, chmod 600 on the keychain fallback file
Run a dry-run test (file exists, syntax valid, --help works)
Phase 3.5 — Self-audit the generated launcher
Why: the user may not have known every cyber requirement. The skill must verify its own output before declaring success.
What: runs ${CLAUDE_SKILL_DIR}/scripts/audit-launcher.py against the rendered tree. 30+ rules (no vendor URLs, no plain secrets, VPN check present, telemetry kill switches all set, etc.). Prints findings ranked P0/P1/P2.
The skill should present findings interactively: "Here are 3 things I noticed — want me to fix them, ask you, or leave as-is?"
Phase 3.6 — URL purge sweep
Defense in depth: a second pass that specifically catches vendor URLs leaked anywhere outside the explicit deny lists. Runs ${CLAUDE_SKILL_DIR}/scripts/url-purge.py.
Phase 3.7 — Pixel-art banner
Generate the launcher's startup banner via ${CLAUDE_SKILL_DIR}/scripts/pixel-art-logo.py. Saved to <install_dir>/banner.txt; the launcher's show_banner() prints it at every launch.
Phase 4 — Generate the distribution kit
Based on the chosen distribution mode:
Public / private git repo — scaffold a clean tree, .gitignore, LICENSE, README.md, with gh repo create --push ready to fire.
Tarball — <slug>-<version>.tar.gz + SHA256SUMS.
One-liner URL — emit a hosted install.sh plus the exact curl ... | bash command to share. Includes a checksum step.
No distribution — skip.
See references/distribution-modes.md for the security caveats of each mode.
Phase 4.5 — Compliance .docx
If the user wants to share the launcher's compliance posture with their RSSI/CISO, generate compliance.docx ready to send. 10-section Word document covering architecture, threat model, cyber controls, network perimeter, audit log, offboarding.
Phase 5 — Post-install
Print:
the exact command the user runs locally (acmecorp-ai, pcode, etc.)
the install path
the uninstall command
the distribution artifact (repo URL, tarball path, or one-liner) the user shares with their team
a checklist of follow-ups (API key not yet set, VPN required, RSSI sign-off pending)
"Proudly made from France with ❤️"
Reference files (read on demand)
references/interview-flow.md — exhaustive question script with the 9 sections
references/provider-matrix.md — which CLI supports which backend
references/env-vars.md — env vars per CLI per backend
references/security-patterns.md — proxy detection, CA bundle, VPN gate, secret storage
references/skills-bundle.md — how the skills bundling works, presets available
references/distribution-modes.md — the 4 distribution modes and their trade-offs
dist/acme-copilot-1.0.0.tar.gz + dist/SHA256SUMS, or
dist/repo/ — git-ready tree, or
dist/install.sh + the one-liner string curl https://acme.internal/install.sh | bash
Quality bar before reporting "done"
Every templated file must render without ${UNRESOLVED} left in the output.
The launcher must --help cleanly without launching the underlying CLI.
A dry-run with <SLUG>_DRY_RUN=1 must print every env var it would set and exit 0.
The uninstall must remove every file the install created, plus the shell RC block (idempotent).
The system prompt rebrand must hide every mention of the underlying CLI vendor and replace with the corporate brand.
The distribution kit must include a checksum or signature when shipping via a one-liner.
If any of these fails, fix it before declaring the install done.
Anti-patterns (and the positive rule that replaces them)
Each rule is phrased as an imperative, with the because that justifies it. If a rule conflicts with the user's environment, surface the conflict in the post-install summary instead of silently breaking it.
Route every request through the corporate gateway (Bedrock, Azure OpenAI, Vertex, LiteLLM) — never call the vendor's public API directly, because the corporate gateway is the only contractually-authorized egress and direct calls bypass DLP, billing attribution, and the legal review the org already signed off on.
Store the API key in the OS keychain, with a chmod 600 dotfile fallback — never plaintext, never world-readable, because chmod 600 + keychain is the documented baseline that the cyber sign-off relies on; a world-readable key is treated as a leaked key.
Scope SSL trust to the process via NODE_EXTRA_CA_CERTS / CODEX_CA_CERTIFICATE / REQUESTS_CA_BUNDLE — never disable SSL verification globally, because the corporate CA must be trusted only by the launcher, not by every other tool on the host; a global toggle is a foot-gun the security team will refuse to ship.
Keep every change process-level and reversible by uninstall.sh — never touch /etc/hosts, the system trust store, or any global config, because the uninstall must restore the machine to its pre-install state without root so endpoint management tools don't flag drift.
Lock the wrapped CLI to a pinned version — never enable the underlying CLI's auto-update inside the launcher, because an unannounced upstream change can break the brand override, the SSE strip-proxy, or the cyber rules; upgrades go through a controlled bump + re-distribution.
Ship the one-liner with a checksum or signature step (SHA256SUMS, cosign, or minisign) — never publish a bare curl ... | bash, because without integrity verification, a compromised host or MITM can swap the installer and every colleague becomes a foothold.
Bundle only skills the user has reviewed (or that come from a vetted internal source) — never pull from a random public repo at install time, because the cyber team needs an audit trail for every prompt, hook, and MCP server that runs on a corporate machine.
