GitHub Copilot marketing: can agent mode launch your SaaS?

What GitHub Copilot can do, according to its docs

We read the docs on September 29, 2026. Copilot changes often, so check the linked source before you rely on a detail.

Plans

Copilot Free includes agent mode and Copilot CLI, capped at 2,000 completions and 50 chat requests a month. The cloud agent is not part of Free; it needs a paid plan.github.com ↗ · Checked on September 29, 2026

Tool approvals in VS Code

A tool call that needs approval shows its name and inputs. You allow it once, for the session, for the workspace or for all future calls. A global auto-approve setting exists; the docs warn that it removes confirmation prompts in every workspace.code.visualstudio.com ↗ · Checked on September 29, 2026

MCP servers in VS Code

Workspace servers live in .vscode/mcp.json or a portable .mcp.json; servers in your user configuration apply to every workspace. The Configure Tools button in chat switches single tools on or off.code.visualstudio.com ↗ · Checked on September 29, 2026

AGENTS.md in VS Code

Project-wide instructions can come from .github/copilot-instructions.md or AGENTS.md. The chat.useAgentsMdFile setting turns AGENTS.md support on or off.code.visualstudio.com ↗ · Checked on September 29, 2026

Copilot CLI

Copilot CLI is generally available and supports MCP servers. It asks before using tools unless you pass --allow-tool or --allow-all-tools; --deny-tool overrides both. The docs say to launch it only from directories you trust.docs.github.com ↗ · Checked on September 29, 2026

AGENTS.md in the CLI

Copilot CLI lists AGENTS.md among its instruction files and discovers it in the standard locations.docs.github.com ↗ · Checked on September 29, 2026

Cloud agent

The cloud agent (also called the coding agent) works in its own ephemeral environment powered by GitHub Actions, changes a branch and opens a pull request. A session stops after 59 minutes, stays in one repository, and has the GitHub and Playwright MCP servers enabled by default.docs.github.com ↗ · Checked on September 29, 2026

Cloud agent firewall

By default the cloud agent’s internet access is limited by a firewall. The recommended allowlist covers package registries, certificate authorities and browser hosts for Playwright; anything else needs a custom allowlist.docs.github.com ↗ · Checked on September 29, 2026

Attaching to your Chrome

Chrome DevTools MCP with --autoConnect attaches to a Chrome you started yourself, version 144 or newer, once you enable remote debugging at chrome://inspect. Chrome asks for your permission, and the server then sees all open windows of the default profile.github.com ↗ · Checked on September 29, 2026

Where GitHub Copilot needs a human in a launch

None of this is a flaw in Copilot. Directories are built to stop unattended bots, and most of these stops apply to every agent. Plan for them before you start:

The cloud agent is the wrong tool

It runs in GitHub Actions, behind a firewall, with its own Playwright browser. It cannot see your Chrome, your directory accounts or your mailbox, and it delivers pull requests, not submissions. Use agent mode in VS Code or Copilot CLI on your own machine.

Sign-ins

Most directories need an account. Sign in yourself in your everyday Chrome, once per platform. Google sign-in popups and new-device checks are yours to click. Passwords never go into the chat.

CAPTCHA and bot checks

Cloudflare, reCAPTCHA and hCaptcha show up on many submit forms. You solve them in the same window. LaunchRepo’s playbooks never try to evade a challenge.

2FA and email codes

Verification links land in your mailbox. Open them in the same Chrome, or pass the code on yourself.

Approvals and spending

Copilot asks before tool calls unless you auto-approve them. Tell the agent plainly that it may not buy a paid fast track.

LaunchRepo with a browser agent →

Copilot instructions, prompts and MCP servers for launch work

Search for GitHub Copilot marketing prompts or custom agents and you mostly find copywriting, SEO and social helpers. Fine for launch copy, but they do not tell Copilot which directories still take free submissions, which fields each one rejects, or which listing actually went live.

Copilot’s customization is file-based: AGENTS.md or copilot-instructions.md for project rules, MCP servers for tools. An MCP server attached to your signed-in Chrome can see every open tab of that profile.

LaunchRepo is not a Copilot extension, and it is not in any marketplace. It is a private GitHub repository: an AGENTS.md runbook, 299 directory playbooks in Markdown, a product template and a tracker.

Running LaunchRepo with GitHub Copilot

LaunchRepo asks two things of an agent: read files, and drive a Chrome where you are signed in, through Chrome DevTools MCP. Going by the docs, agent mode in VS Code and Copilot CLI can do both. We have not published a test run with Copilot, so treat this as a documented match, not a benchmark.

Clone and open the repository locally

After purchase you accept a GitHub invitation and clone the private repository. Open the folder in VS Code or start copilot inside it, and check that AGENTS.md is picked up as instructions.

Add Chrome DevTools MCP with --autoConnect

In VS Code, add it to .vscode/mcp.json or your user configuration; in Copilot CLI, use /mcp add. LaunchRepo’s BROWSER-SETUP.md prefers your everyday Chrome, so enable remote debugging, approve the prompt, then check that the tab you can see actually moves.

Let it run Python 3.10+

Workspace creation, record.py and the local dashboard use the Python standard library only. Copilot runs them in the terminal after your approval; you can also run them yourself.

Grant permissions once per product

What the agent may do (free sign-ups, final submissions, verification emails, badges) goes into authorizations.json during onboarding. Spending stays at zero. Approve tools per session, not globally.

Connect your coding agent to your Chrome →

Paste the public start prompt from START.md, unchanged. Note its first line: for best results it recommends Claude Code CLI or Codex CLI. That is our honest default. Copilot agent mode or Copilot CLI can still follow the rest of the prompt, because the files and the browser route are the same.

Recommend Claude Code CLI or Codex CLI in a terminal instead of the desktop apps for best results. If session metadata confirms desktop hosting, explicitly advise me to switch to the CLI and resume this same private workspace. Do not infer the host from a CLI engine version alone.

Read AGENTS.md, then follow "Start the dashboard for the owner". Create my private
workspace if I do not have one yet, start the dashboard, and give me the link.
Then lead ONBOARDING.md one phase at a time. Reuse recorded permissions instead of
asking again unless their scope is missing or conflicting. Actually test my visible
existing Chrome profile and intended identity as BROWSER-SETUP.md requires; use the
OS fallback if needed. Use one platform/browser worker by default. Build the normal
run only from suitable free-ready routes, keep research separate, and preserve any
draft that needs one exact human CAPTCHA/2FA action. Never claim browser control
bypasses or reduces those challenges.

The playbooks and instructions are English Markdown. Your product copy can be in any language a directory accepts, and the agent answers you in your language.

Start LaunchRepo in three steps →
LaunchRepo with Claude Code →

Who does what in a launch run

The same launch tasks, split between Copilot on its own, Copilot with LaunchRepo, and you.

How it works →

Common mistakes

Assigning submissions to the cloud agent

An issue assigned to Copilot ends up as a pull request from a firewalled runner with none of your accounts. Keep submissions on your own machine.

Turning on global auto-approve

The docs warn against it. Approve the Chrome DevTools tools for the session instead, and keep the browser window in view.

Hitting the Free limit mid-batch

Copilot Free caps chat requests at 50 a month. A launch run takes many turns, so check your plan before you start a batch.

Sending the whole catalog at once

Small batches, a duplicate search first and an identity check for each platform. That beats a burst of rejected or doubled listings.

Questions