Skill MCP

March 27, 2026 · View on GitHub

Last updated: 2026-03-27


v1.0.0-beta — Shipped

Everything below is included in the current release. See CHANGELOG.md for the full feature list.

Core

  • 11 MCP tools: search_skills, get_skill, get_resource, list_categories, install_skill, remove_skill, save_skill, push_skills, update_config, save_config, get_status
  • Hexagonal architecture with port interfaces for all adapters
  • Secure install/save flows with validation + security scanning
  • Trust-aware skill lifecycle and lock tracking
  • Compact manifest injection at MCP initialize time
  • Project-aware config overlays with multi-path discovery
  • Resilience layer: tiered timeouts, rate limiting, circuit breakers
  • Worker manager with heartbeat monitoring
  • Grouped search across local, team catalogs, and GitHub (opt-in)
  • Frecency-aware ranking with local usage analytics (SQLite-backed)
  • Catalog adapters: git clone/pull and static HTTP with cache
  • GitHub code search with opportunistic auth and rate-limit tracking
  • Offline-first: graceful degradation when remotes are unavailable

CLI

  • search with --refresh and interactive remote install picker
  • stats and usage export/reset commands
  • Full --help and --version support

Future Phases

Phase 2: Observability & Security (next up)

TaskDescriptionEffortStatus
Secret scanningThree-layer defense: gitleaks pre-commit + CI workflow + GitHub Secret ScanningSmall✅ Done
Code quality gatesjscpd duplication, eslint-plugin-sonarjs cognitive complexity, gist-backed badgesSmall✅ Done
OXC toolchainoxlint as primary linter (TypeScript + correctness + suspicious), oxfmt as formatter (single quotes), ESLint for security/sonarjs onlySmall✅ Done
Multi-runtime supportBun as default dev runtime, Deno as first-class alternative, mise.toml for pinned tool versionsSmall✅ Done
Editor configsVS Code + Zed configs with OXC formatter/linter, Deno/Bun/Node debug launch configsSmall✅ Done
TypeScript scriptsConverted all scripts/*.mjs → .ts with tsx runner, auto-generated dev-reference.mdSmall✅ Done
OTel implementationWire real spans/metrics into existing Telemetry port; pluggable backends (Langfuse, Jaeger, Grafana) via configSmall
CredentialStore adapterOS keychain + encrypted file backends for API tokensSmall
Enhanced scannerIntegrate Semgrep, ShellCheck, Trivy when available on PATHMedium

See observability research for MLflow/Langfuse analysis.

Phase 3: Transport & Distribution

TaskDescriptionEffort
SSE/HTTP transportSecond transport mode for web-based clientsMedium
IPC socket transportLocal inter-process communicationMedium
Go TUI installerCross-platform rich terminal installer for first-time setupLarge

Phase 4: CLI Completeness

TaskDescriptionEffort
doctorDiagnose config, scanner, connectivity issuesSmall
guideInteractive onboarding walkthroughMedium
manualOffline reference for all tools and configSmall
resetReset config, usage data, or skill stateSmall
uninstallClean removal of skills and configSmall
logsView and tail server logsSmall

Research

TopicSummaryLink
MLflow vs LangfuseNeither is a natural direct integration. OTel foundation first, Langfuse as optional backend, MLflow for separate eval harness.Full analysis

Design Principles

  • Hexagonal architecture — all features go through port interfaces
  • TDD — failing test first, then minimal implementation
  • Offline-first — remote features degrade gracefully
  • Context efficiency — minimize token cost for agents
  • No native deps without justificationbetter-sqlite3 is the only planned native addon