Lab 8 — Skills and MCP tool scaling

Capability: Codex can discover skills and register MCP servers. Skill instructions are not an authorization policy; adding a server exposes tools with the permissions of its process. Tool search and token-tax comparisons require actual measured runs, not just a configured server list.

Install Codex CLI; create a private temporary CODEX_HOME, copy this lab’s config.toml.example to $CODEX_HOME/config.toml and fill in the endpoint (ending in /openai/v1) and deployment. Provide AZURE_OPENAI_API_KEY through your resource key from a secret manager. The Microsoft guide says Entra ID is not supported for Codex; never save the key in files. Workspace: start in lab08-skills-tools/ and follow the common preflight. workspace/ is the with-skill run; workspace-b/ is the without-skill run. Reset both before setup so an old skill cannot contaminate the control. Copy only this lab’s skill, and create matching product fixtures:

$second = Join-Path $labdir 'workspace-b'
$skilldir = Join-Path $workdir '.agents\skills\product-description'
New-Item -ItemType Directory -Path $skilldir -Force | Out-Null
Copy-Item .\product-description\SKILL.md $skilldir
'Demo bowl; material: ceramic; use: food.' |
  Set-Content (Join-Path $workdir 'product.txt') -Encoding utf8
Copy-Item (Join-Path $workdir 'product.txt') (Join-Path $second 'product.txt')
codex exec --skip-git-repo-check -C "$workdir" --sandbox read-only --json `
  "Use the product-description skill to describe product.txt. Cite product.txt."

second="$labdir/workspace-b"
mkdir -p "$workdir/.agents/skills/product-description"
cp product-description/SKILL.md "$workdir/.agents/skills/product-description/"
printf 'Demo bowl; material: ceramic; use: food.\n' > "$workdir/product.txt"
cp "$workdir/product.txt" "$second/product.txt"

Run codex exec --skip-git-repo-check -C "$workdir" --sandbox read-only --json "Use the product-description skill to describe product.txt. Cite product.txt.". Compare a fresh codex exec run with -C "$second" and the prompt "Describe product.txt using only its attributes. Cite product.txt." (without installing a skill there) and check the responses against the actual attributes. Inspect the event stream or Codex’s skill listing in your installed version to confirm the skill was loaded; a prompt alone does not prove it. Alter the skill body in a second disposable copy and compare hashes before trusting a published skill.

Explore MCP safely with codex mcp list and codex mcp add --help. To exercise tool discovery, register only a trusted, learner-controlled MCP server restricted to synthetic data in this lab’s isolated CODEX_HOME; verify codex mcp list --json, run a read-only query, and remove it with codex mcp remove <name>. Do not register production servers or pass a Foundry credential to a third-party MCP process. If no trusted server is available, mark the MCP portion not run rather than claiming a 40-tool comparison. Contrast all-tools vs. a reduced set only when you can inspect actual exposed tools and measure usage in both runs.

For the altered-skill check, finish the no-skill run first, then reset workspace-b/, recreate its product file and copy this lab’s skill there before editing and hashing it. Do not overwrite the original skill fixture. Follow reset and cleanup afterward.