AutoVaultv0.2.1
Markdown
Comparison · v0.2.1 · 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 have felt skill drift — the same skill copied into three projects, each one tweaked, upstream long gone — broad sync alone does not remove the fork. The core AutoVault move is transforms instead of forks: keep pristine upstream source, apply workspace-local deltas at render time, then sign and scope the output each caller actually loads. Skillfish is the closest direct neighbor when you need install, update, sync, and team bundles across many agents. Tessl and SkillKit / Agent Skills are stronger when public discovery is the job. Manual folders are still enough for a few trusted files.

AV
AutoVaultus
Local-first vault. Admits trusted source, then renders scoped variants from pristine upstream plus local deltas.
v0.2.1MITself-hosted
SF
Skillfish
Open-source skill manager for install, update, sync, and team bundles across many agents.
open sourcemulti-agentteam bundles
TS
Tessl
Package and distribution layer for skills and agents, strongest as a published ecosystem.
ecosystemdistributionhosted
SK
SkillKit / Agent Skills
Directory and spec-oriented discovery surfaces for reusable skill source material.
discoveryspecssource material
—
Manual folders
Hand-maintained ~/.claude/skills, ~/.codex/skills, Cursor rules, and repo docs.
zero toolinglocalhigh drift
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
SFSkillfish
TSTessl
SKSkillKit / Agent Skills
—Manual folders
Install, sync, and source of truth
Single source of truth
One canonical SKILL.md vs. agent-specific copies
Yes · Pristine upstream + deltas
Workspace-local overlays render per caller without forking source
Yes · Managed skill source
Strong install/update/sync story, but local tweaks still become copies
Partial · Published artifact
Best when consuming the ecosystem surface
Partial · Spec source
Useful upstream material, not a vault
No · Per-machine drift
Each engineer keeps their own copy
Multi-agent sync
Keep Claude Code, Codex, Cursor, and other hosts aligned
Yes · Scoped profile links
Vault renders only what each profile can load
Yes · Core strength
Broad agent and IDE coverage
Partial · Distribution surface
Depends on the consuming agent
Partial · Examples/specs
Good source shape, sync is external
No · Manual copy
Sync depends on memory and discipline
Team bundles
Share a set of skills across a team
Partial · Remote vault / self-host
Team mode exists, management surface is still narrow
Yes · Strong fit
Skillfish is positioned around team bundles and shared installs
Yes · Distribution model
Good for shared ecosystem packages
Partial · Reference bundles
Usable as inputs, not policy
Risky · High drift
Every machine becomes its own bundle
Local ownership
Can the team keep the working vault local-first?
Yes · ~/.autovault
Plain files, SQLite, signatures, source sidecars
Yes · Local manager
Works well for local multi-agent setups
Partial · Hosted-first posture
Useful ecosystem, less vault-owned
Yes · Plain source
Local after you copy it
Yes · By definition
Local files with no shared policy
Validation & trust
Admission gate
Programmatic checks before a skill becomes usable
Yes · 5 stages
Repair -> denylist -> capability -> dedup -> sign
Partial · Manager checks
Not positioned around a reproducible security gate
Partial · Ecosystem policy
Hosted controls are not the same as local admission
No · Source only
Validation belongs elsewhere
No · No gate
Whatever gets pasted can run
Cryptographic signing
Signatures bind content, source, and gate verdict
Yes · Ed25519 chain
Author -> vault -> mirror
No · Not central
No public positioning around signed gate verdicts
Partial · Trust layer
Treat as ecosystem trust, not AutoVault-style local signatures
No · Not central
Specs do not imply a signer
No · No signing
Trust is social and local
Reproducible verdicts
Same bytes in = same gate verdict out
Yes · Gate v0.2+
Run locally with autovault verify
Partial · Operational checks
Useful manager behavior, but not a signed verdict model
Partial · Hosted review
Not equivalent to local byte-for-byte replay
No · n/a
No admission verdict
No · n/a
No verdict to reproduce
Public denylist
Auditable bad-pattern bundle
Yes · 8 active patterns
Same artifact format as skills
Partial · Safety posture
Not documented as a reusable denylist artifact
Partial · Hosted moderation
Different layer
No · n/a
No gate bundle
No · n/a
No shared scanner
Rendering and scope
Per-caller transformation
Output adapted to the agent's idiom
Yes · Transforms instead of forks
Pristine source stays untouched; local deltas render into signed caller output
Partial · Sync formats
Strong multi-agent reach, not the same as signed render-time deltas
Partial · Agent ecosystem
Depends on the runtime consuming it
No · Spec source
Transformation is left to tooling
No · Manual edits
Each target is its own copy
Scoped delivery
Filter by agent, project, device, and profile
Yes · Four-axis scope
Skill visibility is a vault policy decision
Partial · Profiles/bundles
Good grouping, less emphasis on policy gates
Partial · Access model
Registry access is not per-local-profile scope
No · n/a
No local policy model
No · n/a
Folders are blunt instruments
Progressive disclosure
Agents can search first and fetch full instructions only when needed
Yes · MCP-native
Inventory lookup, exact read, resources on demand
Partial · Manager metadata
Depends on host integration
Partial · Hosted metadata
Good browsing surface, different runtime model
Partial · Spec metadata
Useful if tooling uses it
No · Full file load
Everything is just text in a folder
Operations and openness
Dedup before local use
Stops duplicate-skill sprawl before it reaches profiles
Yes · Admission-time gate
Exact and near-exact checks before signing or profile delivery
Partial · Manager view
Useful post-hoc organization, not a pre-commit admission gate
Partial · Discovery ranking
Catalogs can group similar entries
No · n/a
No local inventory
No · n/a
Duplicates are easy to create
Self-hostable remote mode
Run a private vault behind your VPN
Yes · Docker / Railway
Same gate, your keys, your mirror
Partial · Local/team manager
Good local story; remote vault is not the central shape
Partial · Hosted ecosystem
Use when hosted distribution is desired
No · n/a
No vault service
Partial · File shares
Possible, but policy remains manual
Vendor lock-in risk
If a project disappears tomorrow
Yes · Plain SKILL.md files
Markdown + signed JSON + source sidecars
Yes · Plain skill files
Local installs survive
Partial · Published ecosystem
Artifacts may survive, workflow depends on service
Yes · Plain source
Spec/examples remain portable
Yes · Your files
The drift is yours too
Honesty box

Where AutoVault is genuinely behind. Skillfish currently has broader agent/runtime coverage and a clearer skill manager workflow for install, update, sync, and team bundles. Tessl and SkillKit / Agent Skills ecosystems are better starting points when public discovery or standardization is the main job. Manual folders are still simpler for one person with a handful of trusted skills. AutoVault is intentionally narrower: local-first validation, signing, scoped delivery, transforms instead of forks, remote MCP, and OAuth before a broader management surface.

The wedge

Post-hoc dedup helps you browse a messy list. Admission-time dedup stops a duplicate from becoming local infrastructure. Skill managers can reduce visible clutter after the fact. AutoVault puts dedup in the admission gate, before the skill is signed, scoped, rendered, and made available to a developer profile.

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 need skills to pass a local gate, carry provenance, stay scoped to specific profiles, and render cleanly across more than one agent without creating long-lived forks.

Real signals
  • We have ~/.claude, ~/.codex, and Cursor rules drifting apart
  • Security wants signatures before skills reach developer machines
  • We need to customize a skill locally without losing upstream
  • We want a private vault behind our VPN
SF

Pick Skillfish when…

You mainly need broad multi-agent install, update, sync, and team bundle workflows, and a signed local admission gate is not the primary requirement.

Real signals
  • Agent coverage matters most
  • Team bundles are the core workflow
  • You want a skill manager more than a vault policy layer
  • You are comfortable reviewing trust outside the tool
TS

Pick Tessl or SkillKit / Agent Skills when…

You are looking for public ecosystem discovery, reusable specs, or source material before deciding what belongs in your local vault.

Real signals
  • Discovery matters more than local policy
  • You want examples to adapt
  • You are evaluating agent-skill standards
  • You will still review before local use
—

Skip tooling entirely when…

You have fewer than five trusted skills, your team is one or two people, and manual folders are smaller than the tooling overhead.

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 three big bets

What we believe that the alternatives don't.

Below the feature matrix, AutoVault is shaped by three 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 broad skill managers or discovery surfaces win on convenience.

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.

Bet #3 · Agents will flood the corpus with clones

Dedup belongs at authoring and admission time

Evidence
SkillClone analyzed 20K skills and found 75% of all skills involved in clone pairs, 3.5x ecosystem inflation, only 5,642 unique skill concepts, and 41% of skills in clone families superseded by a strictly better variant.
Bet
As agents author more skills, duplicate prevention has to happen before a draft becomes local infrastructure, not after it is already in every profile list.
If wrong
Post-hoc list cleanup is enough, and admission-time dedup can stay a nice-to-have rather than a core gate stage.
Migrating in

Already using one of the others? Here's the path.

There is no magic importer for every ecosystem. The practical path is to bring source skills or local folders into the vault and let the same gate decide what gets admitted and signed.

From Skillfish-managed skills

Keep Skillfish where it is useful for broad agent coverage, then admit selected local skill folders into AutoVault when you need signing, scope, and transforms instead of forks. Upstream stays clean; workspace-local deltas render at profile sync time.

autovault add-local ./skills/extract-pdf \
  --source skillfish/extract-pdf \
  --sync-profiles
# gate passed · signed · profiles refreshed

From Tessl, SkillKit, or spec repos

Treat external ecosystem entries as source material. Pull the SKILL.md or repo locally, review it, then let the AutoVault gate decide whether it belongs in the vault.

autovault add github:owner/skills/extract-pdf
# fetched source · running gate
# admitted to ~/.autovault with provenance sidecar

From hand-maintained CLAUDE.md / AGENTS.md / .cursorrules

Extract the reusable instructions into one SKILL.md, declare permissions and target agents, then admit that source instead of maintaining separate copies.

mkdir -p ./drafts/extract-pdf
$EDITOR ./drafts/extract-pdf/SKILL.md
autovault add-local ./drafts/extract-pdf --sync-profiles