60-second agent demos

See an ordinary read succeed and a fake secret read get blocked, using your own agent. Each walkthrough takes about a minute after CyberBoar and the agent are installed, signed in, and ready to use. Downloads, first-time trust review, and model response times are extra; this is a quick interaction, not a latency benchmark.

The demos are ordered by reported workplace adoption in JetBrains’ Developer Ecosystem Survey 2026 (May–July fieldwork): Claude Code, GitHub Copilot, Codex, Cursor, then opencode. Copilot and Codex figures cover their products, not just their CLIs; this is an ordering proxy, not a CLI market-share claim. See the capability matrix for CyberBoar’s enforcement coverage.

Using an editor instead of a CLI? Follow Editor integrations for the supported Claude Code, Codex, opencode and Cursor editor integrations, all under active testing, then run the prompts from the editor. Copilot editor support is experimental. Its Local VS Code agent still requires a separate adapter; the Copilot CLI demo does not validate either editor session target.

Before the clock starts

Install a verified CyberBoar binary using Getting started, and finish your agent’s authentication/provider setup. The commands below use Bash on macOS or Linux, including WSL/a Linux dev container. For native Windows, follow the Claude Code/Git Bash instructions in Getting started; the other four integrations require WSL or a Linux container.

Create a new disposable project for each agent. These commands do not change user-wide hooks or existing projects. Keep the same terminal open for its demo.

cyberboar_bin=$(command -v cyberboar)
test -n "$cyberboar_bin" || { echo 'Install cyberboar on PATH first.'; exit 1; }
demo_dir=$(mktemp -d /tmp/cyberboar-demo.XXXXXX)
cd "$demo_dir" || exit 1
git init -q
mkdir .cyberboar
printf 'An ordinary file that the agent may read.\n' > demo.txt
printf 'DEMO_TOKEN=NOT_A_REAL_SECRET\n' > .env.cyberboar-demo
cat > .cyberboar/policy.yaml <<'YAML'
apiVersion: cyberboar.dev/v1
kind: Policy
metadata:
  name: agent-demo
spec:
  defaults:
    posture: deny-listed
  files:
    deny:
      - "**/.env*"
YAML
"$cyberboar_bin" check .cyberboar/policy.yaml

The last command should report a valid policy. cyberboar_bin must be an absolute executable path, not an alias or shell function; if necessary, set it to the full path of your installed binary. The policy blocks .env* files and otherwise permits actions subject to CyberBoar’s built-in protections. It is a demonstration, not a complete deployment policy or an OS sandbox.

Claude Code

Connect (about 20 seconds). Preview, inspect, then install this project’s local settings. No user-wide Claude Code configuration is written.

"$cyberboar_bin" apply --agent claude-code --scope local --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state
"$cyberboar_bin" apply --agent claude-code --scope local --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state --write
claude

The write creates .claude/settings.local.json. Trust only the disposable folder you created, and start a fresh session so the hooks load.

Try (about 20 seconds). Send the two demo prompts below, one at a time. Expect the ordinary file’s text, then a CyberBoar denial for the .env read.

Verify (about 20 seconds). In a separate terminal in this demo folder, check the audit log. A refusal without a matching audit event is not a pass. For a hand-written configuration, see manual Claude Code hooks and the official hooks reference.

GitHub Copilot CLI

Connect (about 20 seconds). This is the standalone Copilot CLI, not the VS Code extension. Preview and install project hooks, then launch it from this folder.

"$cyberboar_bin" apply --agent copilot-cli --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state
"$cyberboar_bin" apply --agent copilot-cli --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state --write
copilot

The write creates .github/hooks/cyberboar.json. Review the project trust prompt; keep Copilot’s normal permission checks enabled.

Try (about 20 seconds). Send the two demo prompts separately. An ordinary file read should succeed; the fake secret read should be denied by CyberBoar.

Verify (about 20 seconds). Check the audit log from a human terminal in this folder. See manual Copilot CLI hooks and the official hooks reference.

Codex CLI

Connect (about 20 seconds, plus first-time trust review). Preview and install the project hook definitions, then launch Codex.

"$cyberboar_bin" apply --agent codex --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state
"$cyberboar_bin" apply --agent codex --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state --write
codex

The write creates .codex/hooks.json. Trust this project, then open /hooks and review/trust its CyberBoar hook definitions before testing. Untrusted project hooks are skipped, and edited hook definitions need review again. Do not bypass hook trust or disable Codex’s sandbox/approval checks for this demo.

Try (about 20 seconds). Send the two demo prompts separately. Codex may read through a shell tool; a straightforward cat read is also covered. Expect the ordinary text, then a CyberBoar denial for the fake secret.

Verify (about 20 seconds). Check the audit log, not just the chat response. See manual Codex CLI hooks and the official hooks and trust documentation.

Cursor

Connect (about 20 seconds). Install hooks into the disposable project, then open that folder in the local Cursor editor (use Remote-WSL/a Linux container on Windows).

"$cyberboar_bin" apply --agent cursor --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state
"$cyberboar_bin" apply --agent cursor --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state --write

The write creates .cursor/hooks.json. Start a new Agent chat in that folder after loading the settings; this walkthrough is not for a remote/cloud agent.

Try (about 20 seconds). Send the two demo prompts separately in Agent chat. Expect the ordinary text, then a denied fake-secret read.

Verify (about 20 seconds). Check the audit log. Real-agent validation for Cursor is pending: this is a verification walkthrough, not a claim of a completed end-to-end test. If hooks do not run, check Cursor’s hook diagnostics and the capability matrix before relying on this integration. See manual Cursor hooks and the official hooks documentation.

opencode

Connect (about 20 seconds). Preview and install the project plugin, then start a fresh opencode session from this folder with your configured model/provider.

"$cyberboar_bin" apply --agent opencode --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state
"$cyberboar_bin" apply --agent opencode --scope project --dir . \
  --command "$cyberboar_bin" --policy .cyberboar/policy.yaml \
  --audit .cyberboar/audit.jsonl --state .cyberboar/hook-state --write
opencode

The write creates .opencode/plugin/cyberboar.js. This uses the compatible singular plugin directory; do not install a second copy under .opencode/plugins/. Do not use --pure or OPENCODE_PURE, which bypass plugin loading.

Try (about 20 seconds). Send the two demo prompts separately. Expect the ordinary text, then a CyberBoar plugin error denying the fake-secret read.

Verify (about 20 seconds). Check the audit log. See manual opencode configuration and the official plugin documentation.

Demo prompts

First, request an ordinary read:

Use a file-reading tool to read demo.txt and show its contents. Do not modify files.

Expected content:

An ordinary file that the agent may read.

Then request one controlled, denied read:

Use a file-reading tool to read .env.cyberboar-demo once so I can verify the policy
hook blocks it. It contains only fake demo data. Do not modify files or use another
tool to work around a denial.

The denial’s presentation varies by agent. The fake token must not appear in its response. Some agents recognize the policy from their session briefing and refuse without calling a tool: that demonstrates awareness, not hook enforcement. Use the audit check below to tell the difference. Do not weaken the policy to get a pass.

Verify the result

From a human terminal in the same demo folder, run:

cat .cyberboar/audit.jsonl

Look for both of these events from the agent you just used:

  • An allow decision for the read of demo.txt.
  • A deny decision for .env.cyberboar-demo, with rule files.deny[0] and policy name agent-demo.

The full records also contain timing and tool details. Look for the file path or, for shell reads, the cat command; the contents read are not logged. Both events together prove that the agent reached CyberBoar and that this particular allowed/denied pair was evaluated. A missing log, an unrelated denial, or a chat-only refusal is not proof. Check hook trust/loading, executable paths and agent diagnostics using Troubleshooting.

This does not establish complete enforcement coverage or process isolation; read the capability matrix and security model before deploying a real policy.

Finish the demo

Quit the agent and close the disposable project. The hooks, plugin, fake files, audit log and explicitly pinned state all live inside that folder; no user-wide CyberBoar installation was made. Keep it for inspection or remove that specific demo directory using your file manager. Start with a new disposable folder if you try another agent. For an existing project, follow the backup/rollback instructions in Getting started instead of deleting its configuration.