Kraken

Spec-Driven Development workbench, powered by Claude.

Kraken is an Electron desktop app for running the SDD loop end-to-end: capture requirements → design the system → break it into trackable tasks → execute with Claude. Built around your local Claude CLI (default) or the Anthropic API.

Why

Most AI coding tools collapse the spec into the chat. Kraken keeps the spec a first-class artifact — versioned markdown that lives next to your code — and uses Claude as the engine that drafts, refines, and executes it. By talking to your installed Claude CLI, your existing subagents, skills, and authentication just work.

What's in the box

• Two backends, one app:
  • Local Claude CLI (default) — spawns claude -p --output-format stream-json. Free if you already pay for Claude Pro/Max.
  • Anthropic API — falls back to the SDK with your sk-ant-… key (encrypted via OS keychain).
• Feature specs: requirements.md → design.md → tasks.md with EARS-formatted acceptance criteria.
• Bugfix specs: bugfix.md → design.md → tasks.md with explicit Unchanged Behavior regression guards.
• Claude Code-compatible agents and skills: reads from the standard .claude/agents/ and .claude/skills/ locations (workspace + ~/.claude/). Your existing subagents and skills load automatically.
• Modern UI: VS Code-inspired activity bar, sidebar views, tabbed markdown editor with preview, dockable streaming chat, status bar. Dark by default.

Quick start

npm install
npm run dev

On first launch:

1. Click Open folder and pick any folder. Kraken creates .kraken/specs/ and .claude/{agents,skills}/ inside it.
2. Make sure the Local Claude backend is active in Settings — it should be by default. If the CLI isn't found:

   npm install -g @anthropic-ai/claude-code
   claude       # log in interactively (one-time)

   Then click Re-detect in Settings.
3. From the Welcome screen or the Specs sidebar, click + New spec and choose Feature or Bugfix.
4. Edit the markdown directly, or click Ask Claude to draft the current section. The CLI runs in your workspace root, so it can read your code.
5. When the section is ready, click Advance phase to move Requirements → Design → Tasks.

Don't want to install the CLI?

Switch the backend to Anthropic API in Settings and add a key:

1. Open https://console.anthropic.com/settings/keys.
2. Sign in (or create an account).
3. Click Create Key, copy the sk-ant-… value.
4. Paste it into Settings. It's encrypted with your OS keychain and only sent to api.anthropic.com.

Pricing is pay-per-token. The local CLI is cheaper for most users.

SDD loop

Idea → Requirements/Bug analysis (EARS) → Design → Tasks (waves) → Implementation
                       ↑                                              │
                       └──────────────────────────────────────────────┘
                          (re-sync when reality contradicts the spec)

Agent library

Click Seed defaults in the Agents or Skills view to install the SDD library into .claude/agents/:

Agent	When to use
spec-requirements-writer	Drafts requirements.md from raw intent. Strict EARS.
spec-design-architect	Turns requirements into design.md with components, data, sequences, testing.
spec-task-planner	Breaks design into dependency waves with verifiable outcomes.
spec-task-executor	Executes one task at a time. Reads spec, edits files, ticks the box.
bug-analyzer	Drives the bugfix Analysis phase. Reproduction, Current, Expected, Unchanged.
codebase-explorer	Read-only grounding before planning. Reports file:line.
code-reviewer	Reviews staged changes for correctness, regressions, reuse, simplification.
test-generator	Maps each EARS statement (and Unchanged Behavior statement) to a test.
spec-doctor	Audits a spec for inconsistencies between requirements, design, and tasks.

Each agent is a markdown file with YAML frontmatter (name, description, optional tools, optional model) — the exact format Claude Code uses. Drop your own into:

• <workspace>/.claude/agents/ (project-scoped)
• ~/.claude/agents/ (global, across all projects)

Workspace agents take priority on name conflicts. Same precedence rules as Claude Code.

In chat, type @ to pick an agent; the picked agent's prompt is injected as the system message.

Skills

Skills follow the open Agent Skills standard: a folder with SKILL.md, plus optional scripts/, references/, assets/.

.claude/skills/sdd-feature/
└── SKILL.md

---
name: sdd-feature
description: Walk through requirements → design → tasks for a new feature.
---

Type / in chat to invoke a skill for the next message. Skills are read from the same workspace/global precedence as agents.

Architecture

• Electron main (electron/main.ts) — IPC handlers for FS, specs, skills, agents, settings; CLI process management; Anthropic SDK streaming.
• CLI integration — spawns claude -p --output-format stream-json --verbose and parses JSONL events. The subprocess inherits the workspace as its working directory, so Claude can read your code and respect your CLAUDE.md. PATH is expanded to include ~/.claude/local/, /opt/homebrew/bin, /usr/local/bin, and /usr/bin for reliable detection inside Electron.
• API integration — @anthropic-ai/sdk's messages.stream() runs in the main process; deltas forwarded over IPC.
• Preload (electron/preload.ts) — typed bridge exposed on window.kraken. Context-isolated, no nodeIntegration.
• Renderer (src/) — React 18 + Tailwind + Zustand.
• Secrets — Anthropic API key encrypted with Electron safeStorage (OS keychain on macOS/Windows/Linux).
• Specs on disk — plain markdown in .kraken/specs/. No proprietary format. Commit them with your code.

File layout in a workspace

your-project/
├── .kraken/
│   └── specs/                       # SDD specs (Kraken-owned)
│       └── user-authentication/
│           ├── spec.json            # phase + metadata
│           ├── requirements.md
│           ├── design.md
│           └── tasks.md
├── .claude/                         # standard Claude Code dirs
│   ├── agents/                      # subagents (workspace scope)
│   │   └── spec-requirements-writer.md
│   └── skills/
│       └── sdd-feature/
│           └── SKILL.md
└── ... your source code ...

Plus the global locations Claude Code already uses: ~/.claude/agents/ and ~/.claude/skills/.

Scripts

npm run dev          # dev server + electron with HMR
npm run build        # production build (out/)
npm run start        # preview the production build
npm run typecheck    # tsc --noEmit for node + web tsconfigs
npm run package:mac  # build a directory bundle for macOS

Website

The marketing/product site lives in website/ — a standalone Vite + React app (React Flow, Framer Motion, Tailwind) with real screenshots of the desktop app. See website/README.md. Run it with cd website && npm install && npm run dev.

License

MIT.