Skip to main content

adk — agents-devkit

Implement, review, investigate, document — composed with an advisor strategy and a self-improving decision-log loop.

Single-repo, multi-agent skill kit. Eight polymorphic skills, nine MCPs, nine subagents, installed via install.sh into Claude Code / Cursor / Codex CLI / Junie at user level.

InstallPhilosophy
Clone + install (Claude Code shown)
git clone https://github.com/sujeet-pro/agents-devkit.git ~/code/agents-devkitcd ~/code/agents-devkit./install.sh --target claude# then in Claude Code:# /adk-setup --init

Features

Eight polymorphic skills

/adk-implement, /adk-review, /adk-investigate, /adk-document, /adk-sync, /adk-setup, /adk-improve, /adk-explain. Each routes internally based on input shape — a Jira URL dispatches the from-jira sub-flow, a PR URL dispatches review-pr, etc.

Multi-agent, one repo

install.sh symlinks the right wrappers into Claude Code, Cursor, Codex CLI, and Junie at user level. No marketplace. Same content, native format per agent.

Question-first, every run

Every skill goes through a mandatory ≤3-question phase before any execution. The Q&A is logged to $ADK_DATA_HOME/improve/learning/decisions.jsonl as training data for the self-improvement loop.

Self-improving by design

/adk-improve reads accumulated decision logs and proposes updates to $ADK_CONFIG_HOME/overrides.yaml. After each run the log rotates so improvements compound over time.

Plan/act tool enforcement

--plan mode literally restricts the implementer to read-only tools; --act unlocks edits. Cline-style separation, not advisor-prose-only.

Deterministic hooks

PreToolUse:Bash safety blocks force-push, hard-reset on protected branches, unrequested PR merges; PostToolUse:Edit validates SKILL.md frontmatter and refuses raw-token writes; SessionStart prints the status banner. Idempotently merged into ~/.claude/settings.json.

Install

adk is distributed as a single repo. Clone it once, then run ./install.sh for each agent you use:

Shell
git clone https://github.com/sujeet-pro/agents-devkit.git ~/code/agents-devkitcd ~/code/agents-devkit./install.sh                       # autodetect installed agents./install.sh --target claude       # one agent./install.sh --target claude,cursor./install.sh --target all          # try every supported agent./install.sh --uninstall           # remove everything by marker; leaves your overrides

install.sh symlinks skills + agents + slash commands into the agent’s config dir, merges MCP configs, and appends a pointer to AGENTS.md in the agent’s global guidelines. git pull propagates updates instantly.

See the full Installation guide for per-agent details and the SETUP.md for env-var requirements.

What this site covers

adk is an opinionated skill kit for a single operator (Principal Engineer): code writing, code review, documentation, and production investigations. It composes scripts (deterministic), MCPs (data access), and an advisor wrapper around every skill so each run is plan → clarify → execute → validate → report.

The reference section is generated from the repository source. To refresh it after editing a skill, agent, MCP, or script, run:

Shell
npm run docs:reference

For local development:

Shell
npm installnpm run docs:dev

New here? Read in this order

  1. Philosophy — operating principles behind every skill.
  2. Installation — clone + install.sh + per-agent wiring.
  3. Multi-agent setup — Claude / Cursor / Codex / Junie capability matrix.
  4. overrides.yaml — the single source of user/company truth (workspaces, repos, data dictionary, defaults).
  5. Concepts — advisor strategy, question-first, decision logs, plan/act, edit-format, hooks.
  6. Reference — generated pages for every skill, agent, MCP, script, and shared file.