Back to changelog

CLI Browser Login and Organization Skills

Sign in to the LLM Gateway CLI from your browser or enterprise SSO instead of typing a password, and pull your organization's shared skills into Claude Code, Codex, OpenCode, Cursor, and other agents with one command. Browser login works on every plan; organization skills are available on the Enterprise plan.

A glowing terminal window linked by a beam of light to a browser window holding a key, surrounded by books, a shield, and robot figures on a circuit board

Signing in to a CLI with a password is a habit worth breaking, and for teams on enterprise SSO it was not even possible: the CLI only knew email and password. Browser login fixes both. The CLI now hands authentication to the dashboard you are already signed in to, including SAML SSO, and gets back its own session.

Approve a Code, Not a Password

1# Opens the dashboard, shows a code in the terminal, waits for approval2npx @llmgateway/cli auth login3
4# Jump straight to your organization's SSO sign-in5npx @llmgateway/cli auth login --sso you@example.com6
7# Remote terminal: print the link and code instead of opening a browser8npx @llmgateway/cli auth login --no-browser --timeout 600

The terminal prints a short code, and the approval page at /connect/device shows the same code next to the account it will sign in. You approve only if they match. The CLI then receives a separate, revocable session, never your browser cookie, and auth logout revokes it on the server. Codes expire, are single use, and are rate limited. For deployments without device authorization, email and password sign-in still works when SSO is not enforced for the email domain, and --key remains the API-key fallback.

Self-hosting? The device flow ships with a migration and the CLI client ID llmgateway-cli; pass --api-url, --dashboard-url, and --gateway-url to point the CLI at a private deployment.

Organization Skills

Enterprise organizations can publish a shared skill library under Organization → Skills: import a single SKILL.md, import a folder with references and scripts, or write a custom skill in the dashboard. Owners and admins publish, edit, disable, and delete; every active member can read; disabled skills disappear from CLI discovery; and each action lands in the audit log without storing the skill content.

Limit Value
Skill name Lowercase, hyphenated, up to 64 characters
SKILL.md Up to 200,000 characters
Supporting files Up to 100 per skill, 1 MB per bundle
Encodings utf-8 text or base64 binary

Developers install them where their agent looks:

1npx @llmgateway/cli skills list2npx @llmgateway/cli skills show code-review3npx @llmgateway/cli skills add code-review --agent claude4npx @llmgateway/cli skills add --all --agent codex5npx @llmgateway/cli skills publish ./SKILL.md --org <org-id>

Targets are agents (the default, .agents/skills/<name>/SKILL.md), codex, claude, opencode, cursor, qwen, and pi; --global writes to the agent's home directory. Installation only writes files and never runs a skill's scripts. Your own tooling can read the same catalog with a project API key: GET /v1/skills lists enabled skills and GET /v1/skills/{name} returns the full bundle.

Fifteen Agents, Plus Your Own

Aider, Qwen Code, and Goose join the launcher, bringing the built-in list to 15 agents that start pre-wired to the gateway with npx @llmgateway/cli launch <agent>. Enterprises can register trusted custom launchers from a JSON definition with agents add, and --gateway-url points any launch at a private gateway endpoint.

Organization skills docs → | CLI docs →