AV
AutoVaultusValidating registry. Author once, render per caller. Local CLI; signed; reproducible.
Comparison · v0.2.0 · 2026-05
The honest breakdown of where AutoVault wins, ties, and doesn't.
The homepage table is a teaser. This page is the long form — including the cases where another approach is genuinely better. We re-evaluate every release; if a competitor closes a gap, we say so here first.
In one paragraph
~30s read
If you publish skills for more than one agent and care about provenance, AutoVault is the only option that solves both with a single SKILL.md and a public, reproducible gate. If you only publish for one agent and don't care about signing, RawHub is fine and free. If your team has been hand-maintaining per-agent forks of the same skill, you're paying the ForkFlow tax — that's the case AutoVault was built to retire.
RH
RawHubPublic skill index. No gate, no signing. Browse and copy-paste from a community wiki.
FF
ForkFlowFork-and-edit catalog. Each agent maintains its own format-specific fork of every skill.
—
ManualOpsNo tooling. Engineers paste skills directly into CLAUDE.md / .cursorrules / AGENTS.md by hand.
Feature matrix
Feature-by-feature.
Twelve dimensions across four sections. Cells show the verdict plus a one-line explanation — never a flat checkmark, because flat checkmarks lie.
Capability
AVAutoVault
RHRawHub
FFForkFlow
—ManualOps
Authoring & format
Single source of truth
One canonical SKILL.md vs. N agent-specific forks
Yes · Single SKILL.md per skill
Transformations rendered per caller
Yes · One per skill
Caller copy-pastes raw
No · N forks per skill
Separate CLAUDE.md, AGENTS.md, .cursorrules
No · Per-machine drift
Each engineer keeps their own version
Per-caller transformation
Output adapted to the agent's idiom
Yes · Manifest-driven
Author declares targets; vault renders and signs each
No · —
Caller is on their own
Partial · Manual fork
Each fork is a separate copy
No · —
Whatever the engineer pasted is what runs
Author-once delivery
Write once and reach every supported agent
Yes · 4 agents from one source
Claude Code, Codex, Cursor, AutoHub
Partial · 1 format per skill
Author picks one; rest is copy work
No · Up to N times
One write per agent fork
No · 1× per machine × N agents
Compounding manual work
Validation & trust
Pre-publish gate
Programmatic checks before admission
Yes · 5 stages, reproducible
Repair → denylist → capability → dedup → sign
No · —
Best-effort community moderation
Partial · Per-fork lint
Inconsistent across forks
No · —
Whatever ships, ships
Cryptographic signing
Ed25519 sigs bind author + content + gate verdict
Yes · Ed25519 chain
Author → vault → mirror
No · —
Trust by URL only
No · —
No signing
No · —
No signing
Reproducible verdicts
Same bytes in = same gate verdict out
Yes · Gate v0.2+
Run locally with autovault verify
No · n/a
No verdict to reproduce
Partial · Best-effort
Depends on fork lint
No · n/a
—
Public denylist
Auditable bad-pattern bundle
Yes · 8 active patterns
Same artifact format as skills
No · —
—
No · —
—
No · —
—
Operations
Dedup at submission
Stops duplicate-skill explosion before it starts
Yes · Text in V1
Embedding in V2 preview
No · Browser ranking only
Near-duplicates surface
No · —
Forks are duplicates by design
No · —
—
Self-hostable
Run a private vault behind your VPN
Yes · Docker / Railway / Render
Same gate, your keys, your mirror
Partial · Mirrors only
No gate to self-host
No · Cloud only
—
Yes · By definition
Each machine is the host
Remote MCP endpoint
Mobile and sandboxed agents fetch over HTTPS
Yes · Bundled MCP server
One-click deploy, signed responses
No · —
Browser-only
Partial · Per-agent
Format-specific endpoints
No · —
—
License clarity
SPDX license declared and surfaced
Yes · Required field
Gate-checked
Partial · Optional
Often blank
Partial · Per-fork inconsistency
—
No · Lost in copy-paste
—
Cost & openness
Pricing
Cost to author, host, or consume
Yes · Free, MIT
All open. No paid plans.
Yes · Free
Donations
Partial · Free tier + paid
Private forks behind paywall
Yes · Free
Labor cost is high
Vendor lock-in risk
If the project disappears tomorrow
Yes · Plain SKILL.md files
Markdown + signed JSON
Yes · Plain text
Caller has copies
Partial · Forks survive
Formats drift
Yes · Yours already
—
Honesty box
Where AutoVault is genuinely behind. RawHub has more skills (3,400 vs. our 241) — community indexes always do. ForkFlow has a richer GUI for browsing per-agent forks; if you live mostly in one agent's tooling and don't care about cross-agent skills, that may matter to you. Hosted AutoVault is intentionally narrow right now: fast launch, remote MCP, OAuth, and signed skill delivery before the broader team-management surface grows.
When to pick what
Pick by situation, not by feature count.
Four short cards with the real-world signals that should push you toward each option. We try not to recommend ourselves when we shouldn't.
AV
Pick AutoVault when…
You ship skills for two or more agents, can't tolerate drift, and need a verifiable provenance chain before anything reaches a developer's machine.
Real signals
- We have CLAUDE.md, .cursorrules, and AGENTS.md, and they keep diverging
- Security blocked a registry because there's no signing
- We need to audit which skills loaded last week
- We want a private vault behind our VPN
RH
Pick RawHub when…
You're a solo developer using one agent, breadth matters more than provenance, and you'd rather copy-paste five skills today than configure anything.
Real signals
- Personal projects, not work product
- Single-agent setup
- You're fine reading every skill before pasting
- Discovery matters more than trust
FF
Pick ForkFlow when…
You're already deep in one agent's ecosystem, you don't write skills, and per-agent UX matters more than a unified source.
Real signals
- You consume skills, never publish them
- Marketplace UI matters more than CLI
- Installing the same skill twice is acceptable
—
Skip tooling entirely when…
You have fewer than five skills total, your team is one or two people, and any registry would be larger than the problem.
Real signals
- The whole skill folder fits in one screenshot
- You can name every skill from memory
- Your agents don't support skill autoloading yet
The two big bets
What we believe that the alternatives don't.
Below the feature matrix, AutoVault is shaped by two opinions about where the agent-skill ecosystem is heading.
Bet #1 · Skills are infrastructure, not content
Treat them like SBOMs, not Stack Overflow answers
Implication
A skill that runs in your agent's tool boundary is code in your trust path. Distributing it without a signature is the equivalent of curl | bash.
Bet
In 18 months, unsigned skills will look tolerated for personal use and blocked at the enterprise edge.
If wrong
Provenance becomes overhead nobody asked for, and RawHub-shaped indexes win on volume.
Bet #2 · The agent format wars are permanent
There will not be one universal skill format
Implication
Each agent's idiom reflects how that agent's controller actually reasons. Lowest-common-denominator output degrades every agent.
Bet
The right abstraction is one canonical SKILL.md plus a transformation manifest, not one universal output.
If wrong
A standard emerges and the manifest layer collapses into plain SKILL.md.
Migrating in
Already using one of the others? Here's the path.
We provide first-class importers for the three common starting points. Migration runs through the same gate as authoring, so what you import is what gets signed.
From RawHub
The CLI walks RawHub's index, downloads each skill, and runs it through the gate before it lands in your vault.
autovault import rawhub --user you
# found 47 skills · running gate…
✓ 41 imported · 5 repaired · 1 rejectedFrom ForkFlow (per-agent forks)
Point the importer at all three forks of the same skill. It merges them into one canonical SKILL.md plus a populated transformation manifest.
autovault import forkflow \
--claude ./forks/extract-pdf.cc \
--codex ./forks/extract-pdf.cx \
--cursor ./forks/extract-pdf.cu
# merged 3 forks → 1 SKILL.md + 3 transformationsFrom hand-maintained CLAUDE.md / AGENTS.md / .cursorrules
The importer parses each file, identifies skill-shaped sections, and proposes one SKILL.md per section.
autovault import manual ./CLAUDE.md
# detected 6 skill-shaped sections
✓ 6 SKILL.md drafts written to ./drafts/
autovault publish ./drafts/*.md