Build, evaluate, and maintain agent skills.
One app, many projects. A project is a directory that contains skills; the skill is the primary object — everything else in the Studio is a view of skills or an action on one. Skills, their eval fixtures, and their run records live as files in your project, git-tracked and cloneable with their evidence.
npx skillmaker-studio startThen add your first project from the sidebar's + — point it at any
directory that holds (or should hold) skills, or create a new one. The
Studio runs once per machine and remembers your projects in
~/.skillmaker-studio/config.json; start works from anywhere.
Or install once and use the skillmaker bin day to day:
npm i -g skillmaker-studio
skillmaker project add . # register a project from the terminal instead
skillmaker startnpm resolves a compiled platform binary (macOS arm64 + x64, Linux x64,
and Windows x64) — no postinstall scripts, no runtime downloads. There is also a
/skillmaker skill for driving the
same CLI from inside Claude Code or Codex.
The left sidebar is the spine: your projects and their skills, plus two global views — a Board of every skill by stage and a Tasks queue of work to do across all of them. The center is the Skill page: the live SKILL.md rendered, reviews and stage moves, evals, and activity. The right panel holds a Files browser over the skill's bundle and a Chat panel for agent sessions. Everything updates live as work lands, in light and dark.
Describe the skill you want; an agent frames it, researches, and drafts it in conversation. Each skill gets its own agent session — claude-code or codex, with a model picker, effort control, and image attachments — working directly in your project. The session is transport; everything the agent decides lands on disk in the skill's files.
Evals start from claims: the things the skill is supposed to get right, grouped by risk family (Input / Reasoning / Output / Adversarial / Chain). Fixtures hang under the claims they probe; runs hang under fixtures. Every claim row shows per-model evidence chips at a pinned version — proven, partial, or gap — and "no fixture yet" is an honest state with a one-click path into the Tasks queue, not a hidden one. Runs record whether the skill was actually invoked and carry their grades separately: coverage and validation never blur. Fixtures are runnable straight from the UI ("Run all fixtures"), or from the CLI.
Everything the UI does rides the same engine as the skillmaker CLI.
Agents are first-class users: an agent working in your repo and a human
working in the Studio see and change the same state.
Early software. The chat panel and eval runs drive a real coding agent, so they require Claude Code or Codex installed and authenticated on your machine. The Windows build is new in 0.6.0 and lightly traveled — please open an issue if anything misbehaves there.
packages/core/ # @skillmaker/core — domain: schemas, journal, fold, machine, index
packages/cli/ # @skillmaker/cli — the skillmaker CLI + server (bin: skillmaker)
packages/viewer/ # @skillmaker/viewer — Astro 5 + React 19 + Tailwind 4 UI
packages/skill/ # the /skillmaker skill for Claude Code / Codex
packages/desktop/ # @skillmaker/desktop — Tauri v2 shell wrapping the CLI (macOS)
packages/docs-site/ # @skillmaker/docs-site — Starlight docs (docs.skillmaker.studio)
packages/marketing-site/ # @skillmaker/marketing-site — public site
npm/ # the skillmaker-studio npm wrapper + platform binary packages
docs/ # product plans and design docs (see docs/README.md)
test/e2e/ # end-to-end tests that spawn the real CLI
Bun-workspaces monorepo; package-local guidance in each package's README.
bun install # also clones the Effect research repo (skip: SKIP_EFFECT_CLONE=1)
bun test packages # unit tests
bun run build:viewer # build the viewer (required before start / e2e)
bun run test:e2e # e2e against the real CLI
bun packages/cli/src/main.ts --helpCI runs typecheck, unit, viewer build, and e2e on every PR.
MIT © SocioTechnica