Home › AI Agent Starter Kit
AI Agent Starter Kit
A tested TypeScript agent: tool-calling loop, provider adapter, MCP server template and evals, with agent rules for every tool.
$39 one-time · v1.1.0 · zip, 219 KB
Buy AI Agent Starter Kit: $39Secure checkout by Lemon Squeezy (merchant of record). One-time payment, instant download.
Or get it in the Complete Vibe Coding Toolkit (all four kits for $99).
What's inside
- A TypeScript tool-calling agent loop with a provider adapter (Anthropic via fetch, plus a mock provider so tests and evals run offline).
- An MCP server template (stdio) with configs already wired for Claude Code (.mcp.json), Cursor and VS Code.
- An eval harness (5 cases) and 11 tests.
npm run checkruns lint, typecheck, tests, build and evals. - Agent-architecture rules, plus /add-tool and /add-eval skills.
- Local rules in
src/toolsandsrc/providers, plus the agent-safe definition-of-done, version-pinning and /record-decisions from the Rules Pack 1.1.
Contains a ready-to-run starter project with every tool's rule files already installed.
Rules (11)
| Rule | When it applies |
|---|---|
agent-architecture | How the agent loop, providers, tools and evals are structured in this starter and how to extend them safely |
agent-workflow | How the AI agent plans, executes and verifies work in this repository |
core | Always-on engineering rules for any language or stack in this repository |
definition-of-done | What "done" means; forbids weakening tests, assertions or acceptance criteria to make work pass |
git-workflow | Branching, commit message and safety rules for git operations |
local-rules | Read and obey module-level AGENTS.md files (auth, billing, migrations) before editing files in those directories |
node | Node.js 20+ server conventions (ESM, errors, config, HTTP, processes) |
security | Secrets handling, input validation, authentication/authorization and dependency hygiene rules |
testing | How to write and run tests (Vitest/Jest for TS, pytest for Python) |
typescript | Strict TypeScript conventions for .ts/.tsx files |
version-pinning | Detect exact framework and library versions from manifests and lockfiles before writing code, and use that version's APIs |
Skills and slash commands (9)
| Command | What it does |
|---|---|
/add-eval | Add or update agent eval cases in evals/cases.ts and run them. Use when changing prompts, tools or providers, or when the user asks to measure agent behaviour. |
/add-tool | Add a new tool to the agent (zod schema, implementation, registry, unit test, eval case, MCP exposure). Use when the user wants the agent or MCP server to gain a new capability. |
/commit-work | Run all checks, then stage and create a Conventional Commits commit for the current work. Only when the user explicitly asks to commit. |
/fix-bug | Diagnose and fix one bug with a regression test, smallest possible change. Use when the user reports a bug, error message, failing test or unexpected behaviour. |
/plan-feature | Plan a single feature before writing code. Writes or updates PLAN.md with goal, files, small steps and done-criteria. Use when starting a new feature or when the user asks for a plan. |
/record-decisions | End-of-session step that appends the decisions made in this session (what, why, alternatives rejected) to the Decisions log in AGENTS.md. Use when the user wraps up a session or asks to record decisions. |
/review-changes | Review the current uncommitted diff or branch for bugs, security issues, missing tests and rule violations before committing. Use when asked to review, check or audit changes. |
/security-audit | Audit the codebase or a diff for secrets, missing input validation, missing authorization, injection, unsafe dependencies and insecure config. Use when asked for a security review or before a release. |
/write-tests | Write focused, deterministic tests for a given module or behaviour following the project's test framework. Use when asked to add tests or increase coverage for specific code. |
Every tool, native format
- Cursor:
.cursor/rules/*.mdcwith description/globs/alwaysApply frontmatter. Workflows ship as manual @rules. - Claude Code: a thin
CLAUDE.mdthat imports the sharedAGENTS.mdwith@AGENTS.md, plus.claude/skills/*/SKILL.mdslash commands. - Codex:
AGENTS.md(well under the 32 KiB default cap) with a Decisions log, and.agents/skills. - Windsurf:
.windsurf/ruleswith trigger frontmatter (each file under 12,000 characters) and.windsurf/workflows. - GitHub Copilot:
.github/copilot-instructions.md,.github/instructions/*.instructions.mdwith applyTo, and.github/prompts/*.prompt.md.
Every file is generated from one master source and checked by a format validator, so the five versions never drift apart. In your repo, AGENTS.md is the one file to edit: Codex, Cursor, Windsurf and Copilot read it, and Claude Code imports it. Why one source of truth.
Install
Unzip, then copy your tool's folder to the project root. See the docs for your tool.