saferskills 0.2.0

Every AI capability, independently scanned. Install Skills and MCP servers to any AI agent with a verified SaferSkills trust score checked at install time.

npx saferskills info mcp-server-github        # see an item's score + findings
npx saferskills install mcp-server-github     # install to your detected agents
npx saferskills capability ./my-skill  # scan a local capability (Skill/Hook/MCP/Plugin/Rules)
npx saferskills capability             # audit every capability installed across your agents (incl. commands, subagents & plugins)
npx saferskills agent                  # behaviorally scan your running agents

No install required — npx saferskills <command> runs the prebuilt native binary. Or install it permanently:

npm install -g saferskills      # npm
cargo install saferskills       # crates.io

What it does

SaferSkills scans Skills, MCP servers, hooks, and plugins for security, supply-chain, maintenance, transparency, and community signals — producing a public, methodology-backed trust score with a full rule trace. The CLI is a thin, fail-open client of the public API: reads are unauthenticated and uncapped, and the safe path is the easy path.

Commands

Global flags

--json (machine-readable output on stdout), --no-color / --color <auto\|always\|never>, -v/--verbose, -q/--quiet, --yes, --force, --non-interactive (alias --no-input).

Output discipline: stdout is machine data (JSON), stderr is everything human (steps, warnings, errors, the banner). Honors NO_COLOR, CLICOLOR_FORCE, and TERM=dumb. The two-line SaferSkills banner prints on every command: a full-width brand rule — sized to the exact terminal width and tinted a fresh, calm tone from a curated palette on each run — above a dimmed v<version> · An OpenLatch project line (suppressed under --json/--quiet, and for completion/man).

Show your score — README badge

Every scanned capability has a live trust badge that re-renders on each re-scan. Embed it in your project's README so installers see the independent SaferSkills score before they run anything:

[![SaferSkills 92/100](https://saferskills.ai/badge/<scan_id>/<score>.svg)](https://saferskills.ai/items/<slug>)

The badge links to the full public report at saferskills.ai/items/<slug> — score, four-tier breakdown, every rule that fired, and the vendor right-of-reply. Copy the exact snippet for your capability from its report page (the ⧉ Copy to MD button on saferskills.ai/items/<slug>).

Configuration

State lives under ~/.saferskills/ (override with SAFERSKILLS_DIR):

• config.toml — api_url, min_score, telemetry.
• installs.json — the install registry, used by install / uninstall / update and by list to show the live current score of a CLI-installed capability.
• scan_cache.json — the local scan-results cache. The no-target capability audit writes each scored capability here (keyed by a content hash of its files, drift-aware) so list can show a score for a capability that was previously scanned but never installed via the CLI. Entries older than 90 days are dropped. The audit does not read installs.json — it audits whatever is installed across your agents' own config dirs (skills, MCP servers, hooks, rules, slash commands, subagents, and the active version of each installed plugin), regardless of how it got there, so you need no prior saferskills installs to audit your setup.

The API origin resolves as SAFERSKILLS_API_URL env → config.toml api_url → https://saferskills.ai.

The install score gate resolves as SAFERSKILLS_MIN_SCORE env → config.toml min_score → 90. An item scoring below it (or unscored) warns and asks before installing; a red-tier (< 40) item requires typing the item name. --yes confirms a below-threshold install; only --force bypasses the red-tier type-name gate.

Telemetry

Two anonymous, privacy-preserving channels — never arguments, names, paths, or any personal data:

• Usage analytics — which command ran, its exit code, a coarse duration. Off by default and asked once: the first interactive run prompts you and saves your answer to ~/.saferskills/config.toml. Force on with SAFERSKILLS_TELEMETRY=1.
• Install reporting — an anonymous agent + capability-kind count when you install something, powering catalog popularity. Sent automatically; no prompt.

Both are silenced together by SAFERSKILLS_NO_TELEMETRY=1 (DO_NOT_TRACK and CI are honored the same way), and non-interactive runs never prompt. See https://saferskills.ai/privacy.

Source and fork builds send nothing on either channel: telemetry requires a key baked in at release time, so any binary you build yourself is always inert.

Building from source

cd cli
cargo build --release          # → target/release/saferskills
cargo test
cargo clippy --all-targets -- -D warnings

TLS is rustls-only (no OpenSSL). The binary is a single ~5 MB executable with sub-second start.

Local development

Run any command straight from the workspace with cargo run — everything after -- is passed to the CLI, exactly as if you'd typed saferskills …:

cd cli
cargo run -- info mcp-server-github          # → saferskills info mcp-server-github
cargo run -- --json info mcp-server-github   # global flags go after `--` too
cargo run -- capability ./my-skill
cargo run -- completion bash

A debug build is slower to start; for a release-speed binary use cargo run --release -- <args>.

Pointing at a local API

Source builds send no telemetry (the analytics key is baked in only at release time), but they still hit the public API at https://saferskills.ai by default. To develop against a backend running locally, override the origin with SAFERSKILLS_API_URL:

bash / zsh — prefix the command (scopes the var to that one invocation):

SAFERSKILLS_API_URL=http://localhost:8000 cargo run -- info mcp-server-github

# or export it for the whole shell session:
export SAFERSKILLS_API_URL=http://localhost:8000
cargo run -- info mcp-server-github

PowerShell — set $env: first (PowerShell has no inline VAR=val cmd form):

$env:SAFERSKILLS_API_URL = "http://localhost:8000"
cargo run -- info mcp-server-github

# clear it again when done:
Remove-Item Env:\SAFERSKILLS_API_URL

Environment variables

Precedence across all config is CLI flags → SAFERSKILLS_* env → config.toml → defaults.

Publishing

Published via npm Trusted Publishers (OIDC) and crates.io Trusted Publishing — no long-lived registry tokens in CI. The release path is .github/workflows/publish-npm.yml: it builds 5 platform binaries, signs them (cosign keyless) with SBOMs, publishes the scoped @openlatch/saferskills-<platform> packages + the unscoped saferskills main package, and the saferskills crate.

Part of SaferSkills — every AI capability, independently scanned. · An OpenLatch project · saferskills.ai