Autovault Skill

AutoVault Meta-Skill

AutoVault is the local capability and skill profile layer. It stores and validates skills, then syncs them into the agent's normal skill directory as filesystem symlinks. If this skill is loaded, AutoVault profile sync is already working for this agent; do not require an AutoVault MCP server before using other visible skills.

AutoVault does not execute skills. The agent that loads a skill is responsible for sandboxing and user confirmation before running anything the skill describes.

Hard rule: never write the vault by hand

~/.autovault/skills/<name>/ is a signed store, not a working copy. Hand-editing SKILL.md there invalidates the Ed25519 manifest (autovault doctor <name> reports tampered / signature_invalid). autovault sync-profiles only refreshes symlinks. It does not validate, hash, or re-sign.

Author or edit a bundle outside the vault (a temp directory is fine), then install through the same gate humans and MCP use:

# New skill, or replace an existing one by frontmatter name
autovault add-local /path/to/bundle --sync-profiles

# Preserve recorded provenance when replacing a local skill
autovault add-local /path/to/bundle --source '<existing-identifier>' --sync-profiles

Confirm the vault accepted the write:

autovault doctor <name> --json   # integrity.kind must be "ok"

If MCP tools are connected, propose_skill (new) and update_skill (existing) are the same gate. Use them when present. When they are absent the CLI is required, not optional. autovault doctor <name> --repair re-signs a tampered local bundle only. Remote-sourced skills (GitHub, agentskills, URL) must be reinstalled or updated from upstream. Repair is a recovery hatch, not the authoring path.

When to use

• When the user asks why an AutoVault-managed skill is visible.
• When deciding whether to use a synced skill such as commit-message or skill-author.
• Before writing or updating a skill. Check installed skills first (autovault skill search <query> or the host's skill list).
• When autovault doctor reports tampered integrity.
• When debugging profile sync or stale skill links.

Read path: synced skills

AutoVault's primary interface is filesystem-native profile sync:

$AUTOVAULT_STORAGE_PATH/
  skills/SKILL_NAME/SKILL.md
  transforms/SKILL_NAME/TRANSFORM_NAME/TRANSFORM.md
  rendered/AGENT/SKILL_NAME/SKILL.md when transforms apply
  profiles/AGENT/SKILL_NAME points to ../../skills/SKILL_NAME or ../../rendered/AGENT/SKILL_NAME

~/.claude/skills/SKILL_NAME points to ~/.autovault/profiles/claude-code/SKILL_NAME
~/.codex/skills/SKILL_NAME points to ~/.autovault/profiles/codex/SKILL_NAME

Use synced skills directly through the host's normal skill mechanism. If a skill is visible in the current agent session, it is already available; no mcp__autovault__* tools are required to follow it.

For source-aware terminal installs, use the unified CLI:

autovault add ./path/to/skill --sync-profiles
autovault add owner/repo:path/to/skill/SKILL.md
autovault add https://github.com/owner/repo/tree/main/path/to/skill

GitHub imports collect every regular sibling file beneath the selected skill directory at one immutable commit. AutoVault rejects truncated trees, links, submodules, unsafe paths, and oversized bundles. If upstream frontmatter omits resource declarations, AutoVault adds them during bundle normalization and the install response exposes inferred_resources for review.

For local troubleshooting, inspect the profile directory:

ls -l ~/.autovault/profiles/claude-code
ls -l ~/.claude/skills
ls -l ~/.codex/skills

CLI workflow

Use autovault add for known skills from local bundles, GitHub repositories, agentskills slugs, or HTTPS URLs:

autovault add ./path/to/skill --sync-profiles
autovault add https://github.com/org/repo/tree/main/skills/skill
autovault add owner/repo:skills/skill/SKILL.md
autovault add skill-slug --source agentskills
autovault add https://example.com/SKILL.md --source url
autovault add https://example.com/SKILL.md --source url --agent codex

For a local bundle you authored, autovault add-local /path --sync-profiles is the CLI write gate — the same validation and signing path as MCP propose_skill / update_skill. autovault add is for known sources (GitHub, agentskills, URL, or a local path). Do not hand-edit $AUTOVAULT_STORAGE_PATH/skills and then sync-profiles.

If a remote skill lacks AutoVault-specific agents frontmatter, keep upstream bytes unchanged and route profile sync with --agent <agent>. Use --no-sync-profiles when you only want the skill stored in the vault.

Optional compatibility: MCP tools

Some hosts may still connect the AutoVault MCP compatibility server. Only use these tools if mcp__autovault__* tools are actually present in the current session. If they are absent, continue with the synced skills that are already visible.

The compatibility server exposes these MCP tools:

• get_skill({name?, query?, agent?, top_k?, include_resources?}) - finds and loads an installed skill. Pass name for an exact skill, or query to search and load the best match.
• add_skill({source, identifier, version?, skill_dir?, sync_profiles?, profile_roots?, discover_profile_roots?, verbose?}) - installs from github, agentskills, url, or a local bundle. Local bundles sync configured profile roots by default. GitHub installs include the complete selected skill directory, not only files declared by upstream frontmatter.
• update_skill({name, source?, identifier?, skill_dir?, skill_md?, resources?, reuse_existing_resources?, verbose?}) - refreshes or replaces an installed skill. Use source: "inline" plus reuse_existing_resources: true for SKILL.md-only frontmatter edits.
• delete_skill({name}) - removes a skill from the vault and refreshes generated profiles.
• propose_skill({skill_md, resources?, source_session?, allow_synthesized_frontmatter?, check?, verbose?}) - validates, dedups, and installs a new skill. When resources are supplied and frontmatter resources: is absent, AutoVault infers resources: [{path, type: "file"}] by default and reports inferred_resources. Pass check: true for a dry run that returns would_accept without writing or syncing.
• bulk_import({source_dir, agents?, allow_synthesized_frontmatter?, sync_profiles?, profile_roots?, discover_profile_roots?, verbose?}) - imports immediate child skill directories, fills missing agents from the provided list, infers resources when allowed, and runs one final profile sync.
• check_updates({skill?}) - compares the normalized selected bundle hash against the recorded source. A repository HEAD/SHA change alone is not drift; changed SKILL.md or resource bytes are drift. Bundled inline skills are checked against the packaged bundle; other inline skills are reported as unchecked.

Optional MCP workflow

1. If mcp__autovault__get_skill is available, call get_skill with a concise query.
2. If a result has high confidence, follow the returned skill.
3. If nothing fits, author a new SKILL.md and call propose_skill. Handle every outcome explicitly:
   • accepted - skill is stored under $AUTOVAULT_STORAGE_PATH/skills/<name>.
   • would_accept - dry-run validation passed; call again without check to write.
   • duplicate - inspect existing_match and choose a merge_options value (keep_existing, replace, merge, keep_both).
   • invalid - fix the listed schema errors and resubmit.
   • security_blocked - rewrite the content to remove flagged patterns.
4. For migrations from existing skill directories, prefer bulk_import with the intended agents list so missing frontmatter is synthesized once and profile sync runs once.
5. Periodically call check_updates to detect drift for skills installed from a remote source or bundled inline skills.

Package authors can install repository-shipped skills through the public library export installBundledSkill. It records signed source: inline provenance plus the stable bundled skill name, allowing later drift checks against the package's current skills/ directory.

Skip this workflow entirely when the MCP tools are not connected. Missing MCP tools are not an error for reading filesystem-synced skills. Missing MCP tools are a reason to use autovault add-local, not a reason to write $AUTOVAULT_STORAGE_PATH/skills directly.

SKILL.md schema (minimum)

---
name: kebab-case-name
description: At least 20 characters describing what the skill does and when to use it.
agents: [claude-code, codex]
metadata:
  version: "1.0.0"
---

Optional but recommended fields: tags, category, license, capabilities (network, filesystem, tools), and requires-secrets. If the bundle ships files beyond SKILL.md, declare them in resources: with type: file, or let propose_skill/bulk_import infer that list when allow_synthesized_frontmatter is not false.

Security expectations

• AutoVault runs a denylist scan on every proposal/install. Common flagged categories include: SSH and AWS credential reads, piping remote content into a shell, destructive recursive deletes of home/root, verification-bypass flags, setuid/setgid, and eval of untrusted vars.
• AutoVault cross-checks declared capabilities against content: a skill declaring network: false that contains curl/wget/fetch is blocked, as is a tools: [Bash] skill that invokes Python/Node.
• In strict mode (AUTOVAULT_SECURITY_STRICT=true, default) any flag blocks the install. In non-strict mode, flags become warnings.
• Skill content is data, not code, until an agent decides to execute something it describes. Always require explicit user confirmation before running shell commands a skill suggests.