codeaholicguy/ai-devkit

AI DevKit – a control‑plane for AI coding agents

What it is – A locally‑run npm package that gives you a single configuration file, a terminal UI, and a set of reusable “skills” to orchestrate many AI‑powered coding assistants (Claude Code, Gemini CLI, Codex CLI, Cursor, GitHub Copilot, etc.). It does not replace the underlying models; it simply makes them behave like a coordinated engineering team.

Why it matters – As developers start using several AI agents in the same repository, the workflow quickly becomes chaotic: each tool has its own rule files, memory, and terminal window. AI DevKit adds a thin “operating system” layer that:

• Keeps all agent settings in one .ai-devkit.json file.
• Shows every running session in a live TUI (ai-devkit agent console).
• Lets you route prompts, logs, or stdin to any agent (ai-devkit agent send).
• Provides a local SQLite knowledge base (@ai-devkit/memory) so agents can look up project conventions without stuffing everything into each prompt.
• Supplies composable skills (dev‑lifecycle, verify, tdd, structured‑debug, security‑review, etc.) that enforce a senior‑engineer workflow.

Core concepts

Concept	What it does
.ai-devkit.json	Single source of truth for all agents, their MCP endpoints, and the skill stack you want to use.
Agent console	Text‑based dashboard that lists, inspects, and interacts with live agents across providers.
agent send	CLI command to push a prompt, stdin, or test output to a specific agent or a named group, optionally waiting for a response.
Memory (@ai-devkit/memory)	Local SQLite store for decisions, conventions, and reusable fixes; agents query it on demand.
Skills	Small plug‑ins that can be chained. The flagship dev‑lifecycle walks an agent through requirements → design → planning → implementation → testing → review. Other skills add verification, test‑driven development, debugging, security checks, documentation, and code simplification.

Typical workflow

1. Run npx ai-devkit@latest init – wizard creates .ai-devkit.json and per‑agent folders (.claude/, .cursor/, …) with generated rules and skill hooks.
2. Start agents (e.g., claude-code start …).
3. Open the console: ai-devkit agent console to see all sessions.
4. Use ai-devkit agent send "run the tests" --id my‑claude --wait to ask an agent to run tests and report back.
5. Store a project convention: ai-devkit memory store --title "API handlers return DTOs" ….
6. When a new feature is needed, invoke the dev‑lifecycle skill; the agent creates docs in docs/ai/requirements, design, planning, etc., and only proceeds to code after each gate (verification, tests, review) passes.

Who should use it – Teams that already rely on one or more AI coding assistants and want a reproducible, auditable process:

• Large codebases where conventions must be remembered across sessions.
• Projects with strict quality gates (tests, security reviews, documentation).
• Developers who juggle multiple agents and need a unified UI and messaging layer.

Installation & quick start

# One‑liner – creates the config and installs the needed pieces
npx ai-devkit@latest init

After the wizard finishes, you can run commands such as:

• ai-devkit agent list
• ai-devkit agent console
• ai-devkit agent send "review this branch" --group reviewers
• ai-devkit memory store …
• ai-devkit skill add <registry> <skill> to pull extra skills from the public registry.

Extensibility – Skills are published by many organizations (Anthropic, Vercel, Supabase, Microsoft, Google, etc.). You can add new ones with ai-devkit skill add … or write your own following the SDK guidelines.

License – MIT, no telemetry, runs entirely locally.

---

All details are taken directly from the repository’s README; no additional features are inferred.