Alex: Welcome back. Today we're setting up accessibility agents across the tools people are most likely to use in this workshop: GitHub Copilot in VS Code, Claude Code, Gemini CLI, Claude Desktop, and Codex CLI.

Jamie: And before anybody worries that an agent is some separate mystery system, can we place it in the normal workflow?

Alex: Yes. An agent is a configured assistant you call while you work, usually from chat or a command line. It can help inspect code, draft changes, explain accessibility issues, or guide a test, but you still decide what to run, what to commit, and what evidence is ready to share.

Jamie: So the agent sits beside Git, GitHub, and VS Code. It doesn't replace the contribution workflow.

Alex: The biggest setup rule is tool currency. Accessibility agents rely on newer chat features, API behavior, and bug fixes, so old versions can fail in confusing ways.

Jamie: When you say newer, are we talking about a vague keep everything updated, or are there actual baselines?

Alex: There are baselines. As of May 2026, VS Code should be 1.113 or later, GitHub Copilot should be current from the VS Code Marketplace, Node.js should be version 18 or higher, Python should be 3.8 or higher, and Git should be 2.20 or higher.

Jamie: And the operating system matters too, right?

Alex: Right. Windows should be Windows 10 build 1909 or later, with PowerShell 5.1 or later for command-line tools. macOS should be 10.15 Catalina or later, and you may need to run xcode-select --install for command-line tools. Linux is centered on Ubuntu 20.04 LTS or later, with Debian, Fedora, and RHEL relatives also supported.

Jamie: What about memory and disk space? Some learners are on older laptops.

Alex: For GitHub Copilot in VS Code, Claude Code, and Claude Desktop, plan on at least 4 GB of RAM. Gemini CLI and Codex CLI can work with 2 GB, though more is always nicer. Disk space ranges from about 500 MB to 2 GB depending on the platform, and network access is required for most workflows, with Codex able to support some local modes.

Jamie: So before installing, know your machine, your account, and your connection.

Alex: Exactly. Copilot in VS Code needs VS Code, the Copilot and Copilot Chat extensions, Node.js, Git, a GitHub account, and internet access. Claude Code needs the Claude Code CLI and a Pro, Max, or Team subscription. Gemini CLI needs the Gemini CLI and a Google AI Studio API key. Claude Desktop needs the app, a Claude subscription, and MCP configuration. Codex CLI needs Node.js and the Codex CLI package.

Jamie: I like checking before fixing. What commands tell me whether the machine is ready?

Alex: For VS Code and Copilot, start with code --version. Then run code --list-extensions piped to grep -i copilot if your shell supports grep. Also check node --version, npm --version, and git --version.

Jamie: If the terminal prints a lot of text, what am I listening for?

Alex: For VS Code, you want 1.113.0 or higher. For extensions, you want entries like github.copilot and Copilot Chat. For Node.js, listen for v18.0.0 or higher, and for Git, v2.20.0 or higher.

Jamie: How do the other platforms check in?

Alex: Claude Code uses claude code --version and claude code whoami, plus node --version. Gemini CLI uses gemini --version, gemini config show, and node --version. Claude Desktop is checked inside the app under Settings, About. Codex CLI uses codex --version, node --version, and npm --version.

Alex: For most workshop learners, the recommended path is GitHub Copilot in VS Code. Install VS Code from code.visualstudio.com, open it once to make sure it starts, and then open the Extensions panel with Ctrl+Shift+X or through View, Extensions.

Jamie: Then search for GitHub Copilot?

Alex: Yes. Install both GitHub Copilot and GitHub Copilot Chat, both by GitHub. When VS Code prompts you, sign in with your GitHub account and allow the authentication flow to complete in the browser.

Jamie: Where do Node.js and the agents repository come in?

Alex: Install Node.js from nodejs.org, using version 18 LTS or later, then verify node --version and npm --version. If you are setting up the local Accessibility Agents catalog, clone https://github.com/Community-Access/accessibility-agents.git into your projects folder, enter that folder, and open it with code .

Jamie: That wording helps. Browsing the living catalog is not the same thing as being forced to fork it.

Alex: Exactly. For setup, cloning the catalog can make local agents discoverable. In VS Code, open Copilot Chat with Ctrl+Alt+I, type @, and listen for agent autocomplete suggestions. Also confirm that the .github/agents folder exists. If you want extra testing tools, you can install @axe-core/cli and pa11y globally for accessibility scanning, and playwright globally for browser testing.

Alex: Claude Code starts from npm install -g @anthropic-ai/claude-code. After that, run claude code auth and follow the browser login flow.

Jamie: And this is where the active Claude subscription matters.

Alex: Right. Once authenticated, clone the Accessibility Agents catalog if your workflow needs local access, move into that folder, and try claude code @daily-briefing morning briefing. Gemini CLI is similar at the command line: install it with npm install -g @google/gemini-cli, get an API key from ai.google.dev, set GEMINI_API_KEY in your environment, and test with gemini --version and gemini whoami.

Jamie: Claude Desktop sounds less command-line centered.

Alex: It is. Download Claude Desktop from claude.ai/download, install it, and open it. Then create or edit .mcp.json in your home folder. On macOS and Linux, that is usually ~/.mcp.json. On Windows, it is usually %USERPROFILE%\.mcp.json. The configuration points an MCP server named accessibility-agents at node, passes the path to the catalog's mcp-server index file, and can set PORT to 3000.

Jamie: Then quit and reopen the app?

Alex: Completely quit Claude Desktop, reopen it, and check the logs if the MCP server does not connect. For Codex CLI, install with npm install -g @codex-cli/core, run codex init, answer the interactive setup questions, and test with codex --version and codex role list. To try agents there, you can search roles with codex role search accessibility and run one with codex run @accessibility-lead.

Jamie: After all that, what proves the installation worked?

Alex: Use a short checklist. Confirm Node.js and Git report the expected versions. Confirm the accessibility-agents folder exists if you cloned it, and that .github/agents is inside it. Confirm typing @ in chat or the CLI shows agent names. Confirm the network can reach GitHub and that API checks like claude code whoami or gemini whoami succeed.

Jamie: And then run one tiny agent call instead of trying a full project.

Alex: Exactly. In Copilot Chat, open chat with Ctrl+Alt+I and type @daily-briefing morning briefing. In Claude Code, run claude code @daily-briefing morning briefing. In Gemini CLI, run gemini @daily-briefing morning briefing. In Claude Desktop, type the same agent request into chat. In Codex CLI, run codex run @daily-briefing morning briefing.

Jamie: If the agent responds with useful information, setup is basically alive.

Alex: Yes, and this is a good time to tune accessibility preferences. In VS Code, make sure screen reader support, Accessible View, keyboard shortcuts, high contrast themes, and terminal accessibility settings match how you work. You can also ask agents to answer in shorter chunks, include file paths before explanations, avoid giant tables, or pause before suggesting a command that changes files.

Jamie: Let's talk about the part everyone hits eventually: something installed, but nothing shows up.

Alex: For Copilot in VS Code, if agents do not appear after typing @, reinstall or update both Copilot extensions from the Marketplace. If you get a cannot find agents error, check that the full catalog was cloned and that .github/agents exists. If chat itself is not working, check GitHub sign-in, subscription or access, network restrictions, and extension status.

Jamie: The appendix also mentions cache and restart steps. When should someone use those?

Alex: If updates and sign-in checks do not help, close VS Code, clear the VS Code cache only as directed by current support guidance, then restart VS Code. A simple restart often refreshes extension state, authentication, and workspace indexing without touching project files.

Jamie: What about command-line tools?

Alex: For Claude Code, check authentication, subscription access, Node.js, and whether the claude command is on your PATH. For Gemini CLI, check the API key, environment variable spelling, network access, and gemini whoami. For Claude Desktop, check the .mcp.json path, JSON syntax, file permissions, the server path, logs, and whether the app was fully restarted. For Codex CLI, check Node.js, npm global install permissions, TOML or role configuration, and the fact that some role behavior may still be experimental.

Jamie: And the universal version of troubleshooting?

Alex: Restart the terminal after installing tools, check firewall or proxy blocks, verify exact versions, and copy exact error messages. When asking for help, include your operating system, tool versions, command you ran, what you expected, what happened instead, and any accessibility setup that affects the workflow. Do not paste API keys, tokens, or private account details.

Alex: Updates are not just housekeeping here. Agent support changes quickly, and accessibility fixes often arrive through editor updates, extension updates, CLI releases, and model provider changes.

Jamie: How often should learners update?

Alex: Before a workshop or capstone session, check VS Code, Copilot extensions, Node.js, and whichever CLI or desktop app you plan to use. In VS Code, use the normal app update flow, then open Extensions and update the GitHub Copilot extensions if updates are offered. For npm-based CLIs, rerun the install command with -g for the package you use, such as the Claude Code, Gemini, Codex, or testing tool packages.

Jamie: But don't create chaos five minutes before presenting.

Alex: Exactly. Update one tool at a time when possible, rerun the version checks, run the daily briefing test again, and keep short notes about what changed. If you are in the middle of a live exercise and your setup already works, wait unless the facilitator asks for an update.

Jamie: If setup still does not work, where should people go without feeling like they failed?

Alex: Start with the Accessibility Agents repository and its system requirements, then use the official documentation for the platform you are installing: VS Code Copilot, Claude Code, Gemini CLI, Claude Desktop, or Codex CLI. Release notes and provider documentation can explain behavior that changed recently. The learning cards for installation and setup are useful as quick reminders for commands, checks, and the minimum information to include in a help request.

Jamie: I like ending there. A working setup is great, but a clear help request is also progress. It means you know what you tried, what changed, and what evidence someone else can use to help you get unstuck.