The foundation: one vault, every agent.
The first release brings a local skill vault, a desktop workspace, and an explicit relay system together. Here’s what ships in 0.1.0.
01
Workspace & onboarding
- Open in Workspace with vault totals, managed relays, connected agents, source update status, recent skills, and quick actions for creation, import, scan, and relay.
- Run workspace checks on demand to inspect local health and compare imported sources. Distinguish unchecked, failed, changed, pinned, and ignored source results; review updates before applying them.
- Guide first setup through a welcome, vault selection, and relay defaults, then open Workspace after the vault is ready. Replay onboarding from Settings.
- Use the new cyan relay logo and mint, turquoise, and deep teal colors across the desktop interface and platform icons.
02
Official & community discovery
- Browse built-in skill sources from Anthropic, OpenAI, and Vercel, plus separately labeled community collections from Matt Pocock, Addy Osmani, Superpowers, UI UX Pro Max, and Composio.
- Explore Awesome Claude Code through its community resource-directory link; it curates links rather than installable SKILL.md packages.
- Support repository-root skills and Composio’s larger collection with a bounded 1,024-entry catalog; retain snapshot review, resource checks, and explicit imports.
- Search bounded source snapshots and review skill instructions, resources, license documents, diagnostics, and conflicts before import.
- Offer Quick add for eligible catalog entries, routing entries that need review through the import preview. Recognize skills already present in the vault.
- Keep publisher browsing explicit and separate from a public registry or marketplace. Source checks contact remote hosts only when requested.
03
Vault & library
- Create a local vault or open an existing one. Canonical skill files remain in a regular folder, readable without SkillRelay.
- Browse skills in grid or list view, search by name and description, and filter between your own skills and imports.
- Create skills with instructions, a description, license information, and agent compatibility settings. Edit your own skills and add text resources.
- Inspect SKILL.md, package files, validation reports, agent compatibility, source information, and version history in the skill viewer.
- Preview removal from the Library and require confirmation. Refuse removal while managed relays remain; retain removed skill data and version history for manual recovery.
04
Imports & sources
- Import skills from local folders, ZIP or .skill archives, and Git repositories, including repository subfolders and references.
- Preview package files, content hashes, compatibility reports, diagnostics, and duplicate-name conflicts before importing.
- Resolve imports explicitly: keep the existing skill, import a new one, replace an eligible unchanged import, or create a fork.
- Scan agent skill directories to discover existing skills and identify duplicates, vault matches, conflicts, and shared visibility.
05
Ownership & provenance
- Preserve the original source location, source kind, and Git revision when available. Imported metadata cannot grant local ownership.
- Keep imported skills read-only in the SkillRelay editor. Fork a skill into My Skills to customize it independently.
- Record a fork’s origin skill, content version, source, and creation time alongside its own history.
06
Versions & updates
- Save immutable, content-addressed version snapshots. Display labels or content hashes without inventing semantic version ordering.
- Check imported sources for updates on demand. Preview added, modified, and removed files, updated instructions, and compatibility diagnostics before applying changes.
- Pin an imported skill or ignore a particular upstream content hash to control update decisions.
- Preview and restore earlier versions while preserving current content for recovery. Accept reviewed changes made outside SkillRelay as a new vault version.
07
Agents & relay
- Add local destination integrations for Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Gemini CLI, and OpenCode.
- Inspect installation evidence, native paths, shared discovery aliases, and agent-specific compatibility notes.
- Choose global or project scope and preview the exact source, target path, relay method, and agents that may discover the destination.
- Create live symlinks where directory-link discovery is documented, or explicitly choose a materialized copy. Use copies for Cursor, Copilot, and OpenCode while link discovery remains unverified.
- Review managed distributions, refresh eligible copies, restore missing relays, remove verified distributions, or detach management while retaining files.
- Select multiple agents in one relay workflow, choose a method per agent, review each target, and receive individual results for created, blocked, or review-required destinations.
- Save usual global agents and methods as relay defaults. Quick relay uses those saved defaults when eligible and pauses for conflicts or additional review; saving defaults alone does not relay anything.
- Offer relay after import and an explicit backup-and-replace review for eligible existing links, preserving the prior link for recovery.
08
Validation & Doctor
- Parse the portable Agent Skills format and report frontmatter, naming, description, resource, and agent-extension diagnostics.
- Report incompatible metadata and environment requirements before relay. Format validation is separate from a security review of skill instructions.
- Inspect vault integrity, version snapshots, skill validation, and managed destinations with local health checks.
- Distinguish changes to the canonical vault from copy drift, broken links, missing targets, blocked distributions, and corrupt metadata.
- Present findings for explicit review and reconciliation without background source checks or automatic repairs.
09
Filesystem safeguards
- Revalidate content, paths, and ownership when applying a preview. Refuse modified or unmanaged targets rather than overwriting them.
- Bound package sizes and tree depth; reject traversal, unsafe links, duplicate archive paths, and unsupported metadata schemas.
- Use vault locking, exclusive creation, recoverable replacements, and distribution journals to make conflicting or interrupted operations visible.
- Preserve removed materialized copies in recovery folders. Do not recursively delete agent target directories.
- Treat imported content as data: do not execute skill scripts, render imported HTML, or load remote preview images. Remote Git imports use restricted HTTPS operations.
10
Desktop, CLI & Git
- Introduce a React desktop workspace with Workspace, Library, My Skills, Agents, Updates, Sources, Health, and Settings, backed by Tauri 2 and a shared Rust core.
- Provide CLI commands for status, list, agents, doctor, and init, plus typed JSON requests and a newline-delimited RPC mode.
- Inspect Git status, initialize a vault repository, and create an explicitly confirmed local commit. No automatic commits or pushes.
- Keep the app local-first, with no app account requirement, telemetry, automatic uploads, registry, or marketplace.
- Organize Settings into Relay defaults, Vault, Git history, and Privacy sections, with agent scanning and a way to replay onboarding.
11
npm / pnpm CLI
- Add the @skillrelay/cli Node launcher for the existing Rust CLI, with version-matched native packages targeting macOS, Windows, and Linux on x64 and ARM64.
- Check the packaged binary’s integrity and forward arguments, input, output, and exit status to the same native engine. The launcher performs no runtime downloads or telemetry.
- Add source build, package, and local installation verification workflows, plus the skills list alias. @skillrelay/cli is published on npm; the Rust crate is still built from the repository.
12
Developer foundation
- Add core workflow tests for imports, parsing, path handling, ownership, relay recovery, update decisions, rollback, drift, and vault locking.
- Add a browser E2E harness that exercises the same Rust operations inside an isolated temporary home.
- Include an MIT license, security policy, contribution guidelines, and code of conduct.
Current limits
- The Rust crate is not published yet. Build the core from source until it lands.
- Agent integrations target local environments. Shared skill folders can be visible to multiple agents; local relays do not install skills in cloud sessions.
- Platform support depends on the build environment. Windows junction creation is not enabled; Linux and Windows readiness is not established by a macOS or browser test.
- A public registry, skill publishing, and a marketplace remain future directions. Built-in official and community source browsing and local skill discovery are already implemented.