Overview

SDTK is a suite of six local-first CLI toolkits that turn AI-assisted software work into governed, auditable, resumable engineering. Instead of trusting a chat thread, every phase of work writes real artifacts into your repository — specs, prototypes, review packets, deployment evidence, wiki pages, and run ledgers — with human approval gates where decisions matter.

SDTK works with Claude Code and Codex as execution runtimes. The runtimes execute; SDTK governs.

The suite at a glance

Toolkit	CLI	Owns	Maturity
SDTK-SPEC	`sdtk-spec`	Discovery, specs, PM → BA → ARCH → DEV → QA planning, 17-file SDLC scaffold	Shipped
SDTK-DESIGN	`sdtk-design`	Screen maps, design briefs, static prototypes, design → code handoff	Shipped
SDTK-CODE	`sdtk-code`	Governed implementation workflow, trust layer, review packets	Shipped
SDTK-OPS	`sdtk-ops`	Deployment, verification, monitoring, incident journeys (skill-driven)	Shipped
SDTK-WIKI	`sdtk-wiki`	Local knowledge graph, docs viewer, kanban board, grounded Ask	Shipped
SDTK-AGENT	`sdtk-agent`	Durable multi-step agent orchestration with human gates	Shipped

Four toolkits (SPEC, DESIGN, CODE, OPS) each own one phase of the idea-to-ship pipeline. Two are cross-cutting: SDTK-WIKI is the memory layer every phase reads from, and SDTK-AGENT is the orchestration layer that chains phases into one durable run.

Standalone: SDTK-BRAIN

SDTK-BRAIN is not part of the six-toolkit suite and is not installed by sdtk-kit — it ships on its own (npm i -g sdtk-brain-kit). The boundary is deliberate:

SDTK-WIKI is the knowledge of one product — it indexes that repo's own docs in place and every SDLC phase reads from it.
SDTK-BRAIN is the knowledge of one person — a personal, cross-project second-brain vault living in its own folder, outside any repo, maintained by your agent.

They share a graph builder and viewer, and nothing else. The second-brain pipeline verbs (ingest/compile/discover/enrich) moved out of sdtk-wiki in wiki 0.10.0 and belong to SDTK-BRAIN now.

What makes SDTK different

A governance kernel, not a chat wrapper. Multi-step work runs against a file-backed ledger with a pure state machine: workflow state, events, evidence, approvals, and reports live on disk in your repo. Close the terminal, come back tomorrow, continue exactly where you stopped. See The Governance Kernel.

Evidence over vibes. A PR ships with a review packet. A deploy closes with verification evidence. An agent run renders a durable report of every task, gate decision, and evidence file. See The Trust Layer.

Human gates where they matter. Phase gates in the spec pipeline, review gates before ship, and human_gate stages in agent workflows that block everything downstream until a person decides.

Local-first. Ledgers, artifacts, graphs, and reports are files in your repository. External model/API/runtime calls happen only when you explicitly invoke your AI runtime (for example /orchestrator in Claude Code, or sdtk-wiki ask on Pro — the free Ask tier retrieves locally and calls nothing).

Free and Pro

The core suite is free and MIT-licensed. A one-time Pro purchase unlocks premium capabilities across all kits on a machine:

Project Intelligence — sdtk-spec project ingest / audit / refresh Pro
Wiki Ask — sdtk-wiki ask. Free: ranked, cited passages from your built graph, returned verbatim. Pro: an answer synthesized from those same sources, or NOT IN SOURCES.
Sleep Execute — sdtk-code sleep authorize → sleep execute: after one attended authorization, a bounded doc-class agent session runs headless within hard ceilings (10 iterations / $10 / 24h TTL). Uses your own provider API key; not SDTK-hosted. Pro

Activate once with sdtk activate --license SDTK-XXXX-YYYY. See Installation.