Use SSLZ Agent

Launch SSLZ Agent locally with GitHub Copilot CLI, Claude Code or Codex.

Run SSLZ Agent on your computer to plan an Azure starting foundation aligned with SSLZ. Use GitHub Copilot CLI (the primary path), Claude Code, or Codex. You describe what you are building, and it explains your options and helps you plan them. It only talks and plans: it does not connect to Azure or change anything unless you approve a specific step.

SSLZ Agent is optional. To deploy the landing zone foundation, use the Quick Start and the Classic Terraform or Bicep workflows; that path is the tested default. The agent is a companion for what comes next, such as planning a workload or a migration on top of a foundation that already exists, or for different needs.

This page provides setup instructions only. It does not receive access to your Azure tenant, subscriptions, billing account, or secrets. You can start without Azure CLI, an Azure subscription, or Azure login. The agent, its planning scripts, safety tests, and documentation are all in the repository’s main branch, which is the default branch.

In short

  1. Install Git, Node.js and GitHub Copilot CLI (or use Claude Code or Codex).
  2. Clone the SSLZ repository.
  3. Start the agent and tell it about your project.

The sections below give the exact commands. Everything under “Reference (optional)” is optional reading.

Install the local prerequisites

You need Git and GitHub Copilot CLI. Sign in to Copilot with a GitHub account that has Copilot access. On Windows, use PowerShell 6 or later. Installation options:

Environment Install Copilot CLI
Windows with WinGet winget install GitHub.Copilot
macOS or Linux with Homebrew brew install --cask copilot-cli
npm on any supported OS, with Node.js 22 or later npm install -g @github/copilot

Node.js 22 or later is also needed to run SSLZ’s local .mjs checks and planners. Check your installation:

git --version
copilot --version
node --version

The initial conversation does not require Azure CLI, an Azure subscription, or Azure login. Local JSON planning uses synthetic inputs and does not need Azure access. The website’s Quick Start is a separate, direct operator deployment path with different prerequisites and approval responsibilities.

Clone the repository

The agent profile and its implementation live together on main. The profile is the agent’s instruction file. Clone the repository into a parent directory where sslz-agent does not already exist, using sslz-agent as the clone folder. Git will refuse to clone over a nonempty directory; do not delete or overwrite an existing checkout to follow this guide.

Run each command separately and stop if any command fails. These commands work in PowerShell and macOS/Linux shells:

git clone --branch main --single-branch https://github.com/ricmmartins/sslz.git sslz-agent
cd sslz-agent
git status --short --branch

The status should show the main branch. Review .github/agents/sslz-agent.agent.md, .github/copilot-instructions.md, and any repository configuration before trusting the folder in Copilot. The profile has read, search, execute, and WebFetch tools. Its guidance covers founders and technology leaders across different organization sizes and hosting environments. The startup planner itself accepts teams of 1-50 engineers; larger teams need broader architecture review instead of forcing their inputs into that schema.

Sign in to Copilot and select the agent

Sign in to GitHub Copilot, independently of Azure:

copilot login
copilot --agent sslz-agent

Follow the GitHub sign-in prompts and review the folder trust prompt before accepting it. In an existing CLI session, /login handles GitHub authentication. You can also start copilot, enter /agent, and select SSLZ Agent from this repository. The CLI ID sslz-agent comes from the profile filename.

To start with the guided opening, run:

copilot --agent sslz-agent --interactive "Start a new founder journey."

Use the interactive form so you can review individual permission requests. Do not use prompt mode as an unattended launcher; if the host cannot request an approval, the agent must report that action as blocked. Approve only the specific tool, command, local path, or official documentation URL needed for the next step. Keep the profile interactive and never launch this profile with --allow-all.

Try an interactive request such as:

I am building a B2B API for EU retailers. It is new on Azure, uses PostgreSQL, and must stay in the EU.
Start with repository reads and planning only. Do not log in to Azure, inspect a tenant, install software,
write files, or run a preview. Ask before any command or external request.

Describe your own workload using sanitized facts. For an open-ended request, the agent summarizes what is known and unknown and names the next safe decision, then asks about facts you have not already provided:

What are you building, where does it run today, and what do you want Azure to do first?

Answer one question at a time. A conversation does not guarantee that a deterministic planner ran; request its exact command and review its output when you need a planner result. GitHub also documents --agent sslz-agent --prompt "..." for programmatic invocation. This walkthrough uses an interactive session so you can answer questions and review tool requests; do not add blanket permissions to make an unattended prompt run.

Use Claude Code or Codex

If you use Claude Code or Codex instead, install it from its own documentation, clone the repository as above, and start it in the cloned folder. Claude Code and Codex use generated adapters of the same instructions. They are not yet validated in those clients, so check each permission prompt, keep approvals on, and never use a bypass mode such as --dangerously-skip-permissions or --dangerously-bypass-approvals-and-sandbox.

For Claude Code:

claude --agent sslz-agent

For Codex:

codex "Read agent/adapters/codex-sslz-agent.md and follow it as SSLZ Agent for this session."

The copies live in .claude/agents/sslz-agent.md and agent/adapters/codex-sslz-agent.md. Do not edit them by hand: they are generated from .github/agents/sslz-agent.agent.md and CI fails if they drift.

Approve live access separately

Only when you choose a specific live read should the agent explain the exact read-only repository command, target, required permissions, and information it will read. Then install Azure CLI if needed and authenticate locally for that approved scope. Review the active account and subscription before allowing the inspection. Azure authentication is not consent to subsequent reads or writes.

Keep Copilot CLI’s normal permission prompts enabled. The custom-agent execute capability is broad; the profile does not create an operating-system shell allowlist. Reject commands outside the documented repository entry points. Never use --allow-all, --allow-all-tools, --allow-all-paths, --allow-all-urls, or --yolo with this walkthrough. Do not paste credentials, tokens, billing records, signing keys, customer data, tenant IDs, or subscription IDs into chat or tracked files.

Local artifact generation and live preview need separate review and approval. Provider registration and deployment remain guarded by their operation-specific contracts; provider apply is disabled in this profile. Migration and dual-cloud execution are disabled. The agent does not automatically deploy a workload, and a successful command does not establish application health.

Reference (optional)

If the agent does not appear

A personal profile with the same ID in ~/.copilot/agents/ overrides the repository profile. On Windows, that folder is $HOME\.copilot\agents. Check for both sslz-agent.agent.md and sslz-agent.md, and any conflicting display name. Resolve a conflict yourself before continuing; these instructions do not rename, remove, or overwrite your profiles. If the agent is missing, confirm the repository root, main branch, profile file, folder trust, and CLI version, then restart the CLI. See GitHub’s custom-agent guidance.

Check local planning without Azure

These optional checks read tracked files and sanitized fixtures. Run them yourself, or approve these exact commands when the agent proposes them:

node tests/sslz-agent.mjs

For PowerShell:

node scripts\startup-workload-plan.mjs plan --input agent\examples\startup-input.json --output json

For macOS/Linux:

node scripts/startup-workload-plan.mjs plan --input agent/examples/startup-input.json --output json

The first command checks static profile and journey contracts. The second runs a local workload planner on a synthetic example and prints JSON. Neither proves readiness, quota, available capacity, startup-credit eligibility, or health in your tenant. The full greenfield journey is a different check that writes local artifacts and requires additional tooling; it is not part of this no-Azure smoke check.

The full journey deletes and recreates .sslz/generated/greenfield-journey and creates and removes the process-specific replay store .sslz/remediation-state/greenfield-${process.pid}. Use an isolated checkout to preserve existing reports, and approve both local write targets before running that command through the agent. Generated .bicepparam and .tfvars files can require separate read-only confirmations. If a confirmation is unavailable, the agent must report that readback as blocked; generation alone does not verify the file contents.

Current documentation (optional)

For up-to-date Azure documentation, you can add the public Microsoft Learn MCP server (HTTP, https://learn.microsoft.com/api/mcp). It requires no Azure login or API key. In Copilot CLI, run /mcp add and follow the MCP setup guide.

To check it, run /mcp and look for the server, then start copilot --agent sslz-agent and ask it to cite a Microsoft Learn page. If it says it used a web fetch instead, the profile or runtime does not expose the Learn tools.

A configured server does not guarantee that its tools are exposed to the selected agent. If Learn MCP is unavailable, the agent falls back to approved fetches of learn.microsoft.com pages, or to local facts only, and says so.

What to expect on a real run

A live test with a new Azure account showed this sequence:

Before you start the agent, set the two subscription IDs in the same terminal that launches copilot. In Bash or WSL, run export SSLZ_PROD_SUBSCRIPTION_ID="<prod-subscription-id>" and export SSLZ_NONPROD_SUBSCRIPTION_ID="<nonprod-subscription-id>", then start the agent from that same terminal.

If the agent is already running, exit it (/exit or Ctrl+C twice), export the variables, and start it again. A new copilot session does not see variables you set in another terminal. To check them without printing the IDs, run echo ${SSLZ_PROD_SUBSCRIPTION_ID:+ok} ${SSLZ_NONPROD_SUBSCRIPTION_ID:+ok} in that shell; it should print ok ok. You can also run the preflight yourself in another terminal and keep the JSON file.

  1. Discovery. About a dozen questions, one at a time. The agent keeps your answers separate from facts it has not verified. It then proposes a plan and asks which IaC tool you prefer.
  2. Draft requirements. It asks before writing .sslz/generated/.../requirements.json (a git-ignored local file). You approve twice: the agent’s own question, and then the CLI’s per-command permission prompt.
  3. Read-only preflight. The agent shows scripts/startup-preflight.mjs inspect but does not see variables from your other terminals. Set the subscription IDs in the same shell that launches copilot (or run the command yourself and save the JSON), so IDs stay out of the chat. A blocked status and exit code 1 mean the report found gaps, not that the script crashed. Every check is reported as blocking, including skipped and unknown ones.
  4. Provider registration. If providers are missing, register them yourself with az provider register. The agent never applies this. Azure may return namespaces in lowercase, such as microsoft.insights. Registration takes a few minutes: check with az provider show -n <namespace> --query registrationState -o tsv and rerun the preflight once it shows Registered. Providers stay registered after a teardown, so you do not need to repeat this. Which providers are needed depends on the profiles you chose, for example Container Apps and PostgreSQL.
  5. Readiness gate. Terraform generation requires current evidence that goes beyond the preflight: a second emergency-capable administrator, security, architecture and Bicep/Terraform parity reviews, billing and benefit confirmation, measured recovery tests, and an eligible regional plan. The agent does not accept “I accept the risk” as a substitute, and there is no exemption for accounts without Startup credits. If you cannot supply this evidence yet, use the agent for planning only and deploy with the CI/CD guide or the Quick Start.

Billing checks may stay unknown. The preflight reads billing properties and Billing Benefits credits for each target subscription. A test or pay-as-you-go account, or a missing billing role, can leave these unknown. That is a statement about what could be read, not proof that credits are missing. Ask Azure Billing Support to confirm the billing account and profile for the exact target subscriptions.

Run it in Bash, WSL, or Cloud Shell. The commands in this guide use Bash syntax. On Windows, use WSL, and run az login inside WSL, because the Windows login is not shared with it.

Limits

This is a local CLI launcher. The GitHub.com custom-agent picker is not a supported launch path for this repository, and a hosted agent has no automatic access to your Azure tenant. See GitHub’s invocation reference and custom-agent configuration reference.

The main branch includes adoption conversation guidance, documentation grounding when the required tools are exposed, local contract checks, and fixture-based planning. Checks on sample data are not evidence that the agent ran the planner, and an external planner invocation and a CLI response do not prove that the agent ran the planner itself. This guide does not establish an autonomous conversation-to-deployment flow or real tenant readiness. Record a remotely available revision when you need repeatable evidence. See the SSLZ Agent capability and evidence guide for links to the SSLZ files that implement or plan each capability.