Agent Kit

A stack-agnostic toolkit for setting up AI agent context in any codebase — frontend, backend, or mobile — so Cursor, Claude Code, OpenCode, and other agents understand your project's structure, conventions, and team workflows.

Agent Kit is not an app framework. It generates the guide files (AGENTS.md, layer docs, rules, MCP placeholders) that AI reads before it starts coding.

What is Agent Kit?

When using AI coding assistants, agents often:

• Put files in the wrong places (inconsistent folder structure)
• Miss project conventions (naming, layering, language idioms)
• Repeat MCP and workflow setup from scratch on every project

Agent Kit fixes this by:

1. Analyzing your existing codebase — works for Next.js, Vite, Express, Nest, FastAPI, Django, Expo, React Native, Flutter, Go, Rust, and more
2. Generating documentation and rules tailored to your project (not generic templates)
3. Copying workflow skills (Jira, PRs, Confluence) and MCP placeholders (Atlassian Rovo, Figma, Context7)

The result: the agent has an accurate map of your project before it starts coding.

What does it do?

| Feature | Description | |---------|-------------| | Universal codebase scan | Detects stack, language, and layered structure via a pattern registry (modules, UI, data-layer, routes, handlers, screens, navigation, shared, config) | | Generate AGENTS.md | Agent entry point — stack, dev/test commands, the detected layer map | | Generate architecture guides | .agents/architecture.md, .agents/code-conventions.md, one file per detected layer under .agents/layers/ | | Generate rules | Rules in .cursor/rules/ (or other harness paths) with globs matching your project's actual layers | | Stack-family rules | Tailored guidance for Next.js, Vite, Express, Fastify, Nest, FastAPI, Django, Expo, React Native, Flutter, Go, Rust | | Workflow skills | Jira ticket, create PR, read Confluence PRD, using project context — copied from templates | | MCP registry | Atlassian Rovo, Figma, Context7 — extensible via mcp/registry.json |

Generated vs copied

| Output | Method | |--------|--------| | AGENTS.md, .agents/*, rules | Generated — adapts to your project structure | | Workflow skills, MCP placeholders | Copied — universal templates |

Supported stacks

| Family | Stack | Detection signal | |--------|-------|------------------| | frontend-web | Vite + React | vite and react deps | | fullstack | Next.js | next dep | | mobile | Expo, React Native, Flutter | expo/react-native deps; pubspec.yaml | | backend | Express, Fastify, NestJS | matching deps | | backend | FastAPI, Django | pyproject.toml / manage.py | | backend | Go, Rust | go.mod / Cargo.toml | | unknown | — | falls back to a flat-layout note in architecture.md |

Installation

Package: @fajaralhakim/agent-kit on npm · Node.js 20+

There are two ways to install, depending on your situation.

A. Existing project (recommended)

No CLI install required. Paste one prompt in your agent chat from the project root.

Cursor

Fetch and follow instructions from https://raw.githubusercontent.com/fajaralhakim/agent-kit/refs/heads/main/docs/install/cursor.md

OpenCode

Fetch and follow instructions from https://raw.githubusercontent.com/fajaralhakim/agent-kit/refs/heads/main/docs/install/opencode.md

Claude Code

Fetch and follow instructions from https://raw.githubusercontent.com/fajaralhakim/agent-kit/refs/heads/main/docs/install/claude-code.md

Other harnesses

| Agent | Prompt | |-------|--------| | Antigravity | Fetch and follow instructions from …/docs/install/antigravity.md | | GitHub Copilot | Fetch and follow instructions from …/docs/install/copilot.md |

Base URL: https://raw.githubusercontent.com/fajaralhakim/agent-kit/refs/heads/main

Full list: docs/install/README.md

What happens after you paste the prompt:

1. The agent fetches the install instructions.
2. The agent analyzes your codebase (stack, language, layers, naming).
3. The agent generates AGENTS.md, .agents/architecture.md, .agents/code-conventions.md, per-layer docs, and rules tailored to your project.
4. The agent copies workflow skills + MCP placeholders (Atlassian, Figma, Context7).
5. You fill in credentials (Atlassian OAuth, FIGMA_API_TOKEN, etc.).

Rules are not copied from templates — they are always generated from the project scan so globs and paths stay accurate.

B. Via CLI (npm)

Published as @fajaralhakim/agent-kit on npm. Requires Node.js 20+.

Run without installing (recommended):

npx @fajaralhakim/agent-kit analyze . --json                  # view scan output
npx @fajaralhakim/agent-kit analyze . --write --harness cursor # write files to disk

Install the core context layer (AGENTS.md.hbs, generic rules, using-project-context skill) into an existing project:

npx @fajaralhakim/agent-kit init .

Or install globally:

npm install -g @fajaralhakim/agent-kit
agent-kit analyze . --write --harness cursor

After init, run analyze --write to overwrite the generic AGENTS.md with one tailored to your project's detected layers.

Also works with Bun: bunx @fajaralhakim/agent-kit …

After installation

1. Review generated files

   • AGENTS.md — read first; confirm the layer map matches your project
   • .agents/architecture.md — layer responsibilities
   • .agents/code-conventions.md — naming, path alias, entry pattern
   • .agents/layers/*.md — one per detected layer
   • .cursor/rules/ (or your harness path) — generated rules

2. Set up MCP

   • Copy .cursor/mcp.json.example → .cursor/mcp.json (Cursor) or equivalent for your harness
   • Complete Atlassian Rovo OAuth on first connect
   • Replace YOUR_FIGMA_TOKEN with a Figma personal access token (or remove if unused)
   • Fill placeholders in .agents/atlassian.md and .agents/figma.md

3. Verify

   npx @fajaralhakim/agent-kit doctor

4. Commit to repo

   • Commit: AGENTS.md, .agents/, harness rules folder, MCP example, skills
   • Do not commit: .cursor/mcp.json (secrets) — add it to .gitignore

MCP Registry

Agent Kit ships with a registry of recommended MCP servers. The generated .agents/mcp-registry.md lists ready-to-merge snippets for your harness.

| Server | What it provides | Endpoint / package | |--------|------------------|--------------------| | Atlassian Rovo | Jira + Confluence | https://mcp.atlassian.com/v1/mcp (OAuth) | | Figma | Design context, components, assets | npx figma-developer-mcp (FIGMA_API_TOKEN) | | Context7 | Library + framework documentation | https://mcp.context7.com/mcp |

Source of truth: mcp/registry.json. Add your own MCPs there — the analyzer picks them up automatically.

Workflow skills for Figma (component generation from designs) and Confluence (PRD → tasks) are on the roadmap — they layer on top of the MCP integrations above.

CLI commands

| Command | Purpose | |---------|---------| | agent-kit init [path] | Install core context (generic AGENTS.md, rules, skills, MCP placeholders) | | agent-kit analyze . --json | Scan project → JSON output (ProjectProfile) | | agent-kit analyze . --write --harness cursor | Generate and write project-tailored AGENTS.md, .agents/, rules | | agent-kit add context7 | Add Context7 MCP entry | | agent-kit add caveman | Optional: Caveman guide + lean conversation memory (.agents/memory/) | | agent-kit init . --addon caveman | Core + context7 + caveman addon together | | agent-kit doctor | Validate installation |

Optional addons

| Addon | Command | What it adds | |-------|---------|--------------| | context7 | agent-kit add context7 | Context7 MCP entry in .cursor/mcp.json | | caveman | agent-kit add caveman | Caveman install guide (.agents/caveman.md), lean plan/session memory under .agents/memory/, conversation-memory skill and rule |

Caveman itself is not installed automatically — the addon adds repo-local guidance and memory conventions. Run Caveman's installer from .agents/caveman.md when you want token compression.

Memory flow: agents read compact .agents/memory/index.md first, then only the linked active plan or handoff file — not full chat history.

npx @fajaralhakim/agent-kit init . --addon caveman
npx @fajaralhakim/agent-kit add caveman .

Development (contributors)

Clone this repo to work on Agent Kit itself:

git clone https://github.com/fajaralhakim/agent-kit.git
cd agent-kit
bun install
bun run build

Run the CLI locally:

bun run dev -- analyze /path/to/any/project --json
bun run dev -- analyze /path/to/any/project --write --harness cursor
bun run dev -- doctor /path/to/any/project

Repository structure

agent-kit/
├── docs/install/       # Install prompts per harness (Cursor, OpenCode, etc.)
├── mcp/
│   └── registry.json   # MCP server registry (source of truth)
├── packages/
│   ├── cli/            # @agent-kit/cli
│   └── analyzer/       # Scan codebase + generate docs/rules
├── harness/            # MCP examples per harness (no rule templates)
├── skills/             # Workflow skill templates (Jira, PR, Confluence, project context)
└── profiles/
    ├── _core/          # Generic AGENTS.md + core rules + skills
    └── addons/         # Optional addons (context7, caveman, ...)

License

MIT — free for personal and commercial use. See docs/commercial-use.md.