Skip to main content

Skill Library

The Skill Library holds reusable SKILL.md workflows. A skill describes how to do a task; installing it does not by itself authorize every action it describes. Availability of community operations depends on the configured registry.

Browse, install and use​

  • browse_skills, featured_skills, trending_skills and popular_skills help discover available entries.
  • install_skill installs an available version after package checks.
  • skill_view opens instructions for a selected skill.
  • publish_skill can publish a supported local package after the required checks and authorization. Review every file before publishing.
  • Supported installed skills can be rated, reported, updated or forked.
  • suggest_skills finds relevant available skills.

The Navigation Drawer provides Skill Library. The Skills & Automation panel includes supported skill, learning and task controls. A skill named in a channel request is used only where its instructions, tools and host capabilities are available. An unavailable runner reports a failure rather than silently granted local execution.

Trust before installation​

Skills can contain executable support files or inline shell snippets. Viewing a full local skill can execute its authored snippets when local process execution is enabled. Install only content you trust. A skill description, rating or successful package check is not a security review of everything it can do.

Keep credentials out of skill text and support files. Reference required capabilities or configured connections instead. A skill's instructions do not replace the user's current request, permissions or system approval controls.

Local skill folders​

Operators can configure workspace skills with VRAI_SKILLS_WORKSPACE_DIR, user skills with VRAI_SKILLS_USER_DIR, and additional folders with VRAI_SKILLS_EXTRA_DIRS. Separate extra folders with semicolons on Windows or colons on POSIX. Workspace skills take precedence over same-named user skills.

A connected Web dashboard uses the host's catalog. It does not gain arbitrary access to files on the browser's computer. Discovery and file operations remain subject to platform support and the host's access scope.

Learning and updates​

Automatic memory review is enabled by default; automatic skill review is off. You can request Update memory and skills for an explicit review. Automatic memory writes require citations to direct user, tool or file records from the reviewed session. Unknown, withheld, cross-session, assistant-only and derived entries do not authorize a write. The host checks citations again when applying the proposal. Citations attribute the remembered text to reviewed records; they do not establish that its claims are true.

Repeated memory applications are idempotent, and exact duplicate memories do not use a write slot. A bounded cycle budget counts only new memory and skill mutations. Learning can be paused independently of explicit memory requests. Provenance can be inspected without exposing the cited session text.

If an operator enables automatic skill review, it can save limited notes under an existing skill's references. It cannot automatically rewrite the main skill, create a new one, remove files or change its settings through that review mode.

Operators can control automatic review with VRAI_AGENT_SELF_IMPROVE_SKILL_NUDGE_ITERATIONS and VRAI_AGENT_SELF_IMPROVE_ENABLED. Review proposed notes before relying on them. A best-effort local history can assist recovery, but it is not a guaranteed backup. Keep your own copies before important edits.

Channel approvals​

Configured channel work can run ordinary commands and file edits immediately. The current default channel policy is ask=off with security=full; do not assume it prompts before every potentially sensitive operation. The host account's permissions remain important.

Catalog commands marked approval: required or requires_approval: true normally ask before executing. Catalog read-only markings, explicit approval-option rules and the selected policy affect when a prompt appears. Operators must review those declarations carefully, especially for commands that send messages, publish, delete, make payments or access credentials. A metadata label alone is not a content-safety guarantee.

When a prompt appears, review the skill, command and arguments. Approval applies to that requested action; do not resend it just because a reply is delayed. Changes to server policy may take time to become visible to active sessions. Commands that run without a prompt still use the applicable audit policy.

To restore broader command prompts, operators can use VRAI_ACTION_POLICY_PROFILE=default and VRAI_CHANNEL_EXEC_AUTO_APPROVE=0. Local-app approval rules and separate plugin, memory or protected-file controls remain distinct from channel execution policy.

Server execution and failures​

Server-hosted helper execution requires a compatible build and explicit VRAI_ENABLE_PROCESS_EXEC=1 opt-in. VRAI_DISABLE_PROCESS_EXEC=1 disables it. Review configured monitors as well as skill helpers before enabling execution: background work may send alerts through configured channels.

Use an explicitly intended local host for local shell work. A configured sandbox must not silently change into unrestricted local execution. Browser and unsupported mobile clients do not acquire local process execution from a server setting.

If a skill fails, inspect its reason and available output. Follow another route only when the skill and normal permissions allow it. A retry must not repeat a possibly completed external action without checking the original outcome.