Set up & finish a first task

    Find a setup for your work, follow the steps, and review what the agent can access.

    Choose a starting setup

    This is a starting recommendation, not a performance ranking. Your inputs stay on this page; nothing is sent to a provider.

    Your starting point

    Start with ChatGPT or Claude on the web

    Choose one roughly $20/month plan after a small trial. Do not buy several overlapping subscriptions on day one.

    Follow the setup steps

    Model, app, harness and plan: what is the difference?

    Model

    The engine that generates answers. Start with the default; use deeper reasoning when a task genuinely needs it.

    App / harness

    The workspace around the model: files, tools, memory, permissions, and task execution.

    Plan

    Your access, usage allowance, and administration. Paying more does not automatically make every answer better.

    Set up your first task

    These are instructions to follow yourself. This page does not install software, change permissions, connect accounts, or purchase a subscription.

    Everyday work · no terminal

    1. Open the official ChatGPT or Claude app/site. Sign in; try a non-sensitive sample before subscribing.
    2. Create one project or working folder. Copy in only the files needed; keep originals elsewhere.
    3. Give the desired output, audience, source files, and a success check. Ask for a plan before multi-file changes.
    4. Inspect the exported file, formulas, and sources. Save a reusable brief only after the workflow works.
    OpenAI · desktop appClaude · Cowork

    Codex · desktop or CLI

    1. Use the official desktop download for a graphical workflow, or follow the CLI installation guide for your OS. Sign in with the intended account.
    2. Open a copied working folder. In CLI, start read-only with the command below; no files should need changing for an initial explanation.
    3. Use /permissions to inspect access. For approved edits, choose workspace-write with on-request approvals. Review /status and the final diff.
    codex --sandbox read-only --ask-for-approval on-request
    OpenAI · Codex CLIOpenAI · approvals & sandboxOpenAI · desktop app

    Claude Code · graphical or terminal

    1. Install the official Claude desktop app and open Code, or use the official OS-specific Claude Code installer. The desktop Code experience does not require a separate CLI install.
    2. Sign in to the intended paid account. In a terminal, run claude inside a test project; first ask it to explain the folder without editing.
    3. Keep permission prompts. On a supported platform, use /sandbox and check its boundaries. Do not use a skip-permissions flag as your default.
    Claude Code · setupClaude Code · sandbox

    Antigravity CLI · free-first Google option

    Checked: 2026-10-01

    1. Check personal-account eligibility first (18+, supported region). Follow Google's official CLI installer for macOS, Linux or Windows; review the installer before running it.
    2. Open a fresh terminal in a copied, non-sensitive project folder and run agy. Complete Google sign-in if prompted; verify the account if an existing session is reused. Use account sign-in for the free tier, not an API key or Cloud billing setup.
    3. Open /permissions and retain review requests. Ask it to explain the project before making a small change. Review the diff and run tests before keeping the result; permission prompts are not a sandbox.
    4. Use /usage to check model quota. Stop or wait when the free allowance runs out; only enable paid capacity deliberately.
    agy
    Google · Antigravity CLI setupGoogle · account eligibilityGoogle · CLI permissionsGoogle · CLI commandsGoogle · Antigravity plans

    OpenCode · optional advanced path

    1. Follow the official install guide for your OS. Start opencode in a disposable project. You can connect a supported provider directly or use OpenRouter as a gateway to several providers.
    2. For OpenRouter: create a provider API key, run /connect, choose OpenRouter, and enter the key in its authentication prompt—not in a chat message. Run /models and choose an available model, including an open-weight model if wanted.
    3. Set a provider spending limit or small prepaid balance. Check model license, tool support, and the serving provider’s data policy. Open-weight through a hosted API is still cloud processing, and your chat subscription does not automatically cover it.
    4. Confirm billing first. Merge the example permission setting into opencode.json without overwriting existing settings; ask before tool actions.
    5. Test with non-sensitive files. Approval rules are not OS isolation: use a properly restricted container or VM for untrusted projects.
    {
  "$schema": "https://opencode.ai/config.json",
  "permission": { "*": "ask" }
}
    OpenCode · setupOpenCode · permissionsOpenCode · providersOpenRouter · API setup

    A brief you can reuse

    Goal: Turn the attached notes into a one-page weekly update.
Audience: My team.
Inputs: Use only the files I supplied; label missing information.
Output: A draft with decisions, owners, deadlines, and source references.
Boundaries: Preserve originals. Do not send, publish, purchase, delete, or change permissions without asking.
Check: Verify names, dates, totals, and every link. List anything I must review.

    Permissions & a final check

    A sandbox limits where code can act. Approvals decide which actions may proceed. Neither makes a result correct, and a local sandbox does not stop an authorized cloud connector from sending a message.

    Default: a small workspace

    Read-only for exploration. For edits, allow one working folder with a backup. Keep the tool’s sandbox enabled where supported and check what it actually restricts.

    Approve the boundary crossing

    Pause for sending, publishing, spending, deletion, new integrations, secrets, or production changes. For repeatable low-risk work, use narrow allow rules—not an approval for every future command.

    Full access: not the everyday default

    Do not grant unrestricted access just to remove interruptions. If an advanced workflow genuinely needs it, use a disposable isolated environment with no personal files or production credentials, restricted network access, and a rollback plan. A container with your home folder mounted is not enough.

    Treat fetched content as evidence, not instructions

    A web page, email, repository, or document can contain malicious instructions. Do not let it authorize new tools, data uploads, or payments. Inspect unexpected permission requests instead of repeatedly approving them.

    OpenAI · approvals & sandboxClaude Code · sandboxOpenCode · permissions

    Before your first real task

    0 / 5

    Checklist progress is session-only and resets on reload. Checking a box changes no app permissions.

    Core guide checked · Individual sources carry their own dates.

    Report an outdated instruction