Skip to content

Starter kits

Last verified: 2026-05-06 · Drift risk: low–medium

The repository ships nine starter kits under starter-kits/ on GitHub. Each kit is a copy-pastable project that gives you a working starting point for one platform or pattern.

What every kit contains

File Purpose
README.md What the kit is, who it's for, how to use it.
AGENT_SPEC.md A filled-in agent spec following the Agent spec template.
PROMPT.md The system / instructions prompt used by the agent.
TOOL_ALLOWLIST.md The least-privilege list of tools/connectors the agent may use.
EVALS.jsonl Golden eval cases the agent must pass.
RED_TEAM_CASES.jsonl Adversarial cases the agent must refuse or handle safely.
LAUNCH_CHECKLIST.md Pre-launch checklist tailored to this kit's risk profile.
INCIDENT_RESPONSE.md What to do when something goes wrong.
Optional inert example script Gated behind RIZWAN_APPROVED_TO_RUN=1 so it never runs by accident.

Kits

Kit Best for
universal-agent-spec/ A vendor-neutral baseline. Copy this when starting any new agent.
claude-code-agent/ A repo-aware coding/refactor agent run in Claude Code.
codex-agent/ A repo-aware coding agent run in OpenAI Codex CLI.
gemini-antigravity-agent/ A project-scoped agent for Google Antigravity.
openai-agents-sdk-python/ A Python service that uses the OpenAI Agents SDK with one tool.
openai-agents-sdk-typescript/ The same shape, in TypeScript.
mcp-server-python/ A minimal MCP server using the official Python SDK.
mcp-server-typescript/ A minimal MCP server using the official TypeScript SDK.
local-script-agent/ A local Python script agent with manual-only scheduler examples.

How to use a kit

  1. Browse the kit folder on GitHub.
  2. Copy the files you need into your own repo (or fork this one).
  3. Read AGENT_SPEC.md and edit it to fit your job statement.
  4. Trim TOOL_ALLOWLIST.md to least-privilege.
  5. Run the agent manually at least 3 times before considering any automation.
  6. Walk through LAUNCH_CHECKLIST.md before exposing the agent to anyone else.

Why scripts are inert by default

Per ADR 0003, every example script in a kit:

  • requires RIZWAN_APPROVED_TO_RUN=1 (or an equivalent explicit env var) to do anything,
  • prints what it would do before doing it,
  • and writes a manual-only example for any cron / launchd / Task Scheduler entry — never an enabled one.

This keeps "I copy-pasted a tutorial" from turning into "I have a recurring agent quietly running against my accounts."