Getting Started
Five minutes from clone to first skill invocation.
Fast path
# 1. Clone the repogit clone https://github.com/sujeet-pro/agents-devkit.git ~/code/agents-devkitcd ~/code/agents-devkit# 2. Install for your agent(s)./install.sh # autodetect everything installed# OR explicit:./install.sh --target claude# 3. Set env vars in ~/.zshenv (see SETUP.md §2)export GITHUB_TOKEN_CRED="github_pat_..."export DATADOG_API_KEY_CRED="..."export DATADOG_APP_KEY_CRED="..."export STATSIG_CONSOLE_API_KEY_CRED="console-..."export ATLASSIAN_SITE="acme.atlassian.net"export ATLASSIAN_USERNAME="you@acme.com"export ATLASSIAN_API_TOKEN_CRED="ATATT..."# 4. Restart your agent (Claude Code etc.) so env vars + MCPs loadThen in your agent:
/adk-setup --init # scaffold $ADK_CONFIG_HOME/overrides.yaml/adk-setup --check # verify env, MCPs, agents/adk-setup --enrich # auto-populate metadata cache from MCPsWhat just happened
install.sh did the following for each detected agent:
| Agent | What landed |
|---|---|
| Claude Code | symlinks for 8 skills + 9 subagents + 8 slash commands; MCP merge into ~/.claude/settings.json; 3 hooks (PreToolUse/PostToolUse/SessionStart); CLAUDE.md pointer to AGENTS.md |
| Cursor | 8 rule files in ~/.cursor/rules/; MCP merge into ~/.cursor/mcp.json; global “always” rule pointing at AGENTS.md |
| Codex CLI | 8 prompts in ~/.codex/prompts/; MCP config in ~/.codex/config.toml; instructions pointer to AGENTS.md |
| Junie | ~/.junie/guidelines.md append with AGENTS.md pointer; honest gap doc — Junie’s skill auto-routing isn’t full yet |
All installs are idempotent — re-running install.sh updates symlinks and merge blocks without duplicating. Uninstall via ./install.sh --uninstall --target <agent>.
Try a skill
# Polymorphic on input — works with a Jira URL, GH issue URL, Slack thread, freeform description/adk-implement <Jira URL or KEY-NUM># Polymorphic on target — PR URL, local working tree, doc path, comment thread/adk-review <PR URL># Read-only multi-source investigator/adk-investigate "checkout 500s since 13:00"# Generate any markdown artifact (runbook / ADR / RCA / PR body / commit msg / diagram / ...)/adk-document <intent> --type runbook# Push markdown to Confluence / Jira / GDoc / Slack/adk-sync --write .temp/foo.md --to confluence --target "My Page"Every skill starts with up to 3 questions about scope, constraints, and (when relevant) scale verification. Your answers are training data for /adk-improve.
Next
- Installation — detailed clone + install paths
- Multi-agent setup — per-agent capability matrix
- overrides.yaml — the single config file you maintain
- Philosophy — why adk works this way