Claude CLI Bypass Mode

Bypass Mode passes --dangerously-skip-permissions to your locally installed claude CLI. Every tool call — read, write, edit, run_command, MCP — executes without an approval prompt, fully autonomous. Inference is billed to your Claude Max subscription via OAuth, not to a Hypex API key.

Why use it?

• Free inference — your Max plan covers it. No per-token Hypex Managed credits, no BYOK costs.
• End-to-end autonomy — agent runs without clicking "Approve" on every file edit.
• Dev-tuned system prompt — Claude Code's CLI is built for developers. Combined with Hypex's Systems / Kernel-mode / Research / Game research agent profiles, the model handles every legitimate professional workflow without the false refusals you sometimes see on vanilla API (your-own reverse engineering, security research PoCs, single-player game modding, scrapers on public data).
• Local memory + tools sync — Hypex spawns the CLI in your workspace folder, so it inherits your ~/.claude/CLAUDE.md, custom skills, MCP servers, and project-level steering files automatically.

Prerequisites

1. Install Claude Code CLI: npm install -g @anthropic-ai/claude-code (Hypex offers a one-click "Install via npm" terminal flow when you switch to the claude-cli provider).
2. Authenticate once: run claude in a terminal and follow the OAuth flow. Logs you into your Claude Max account.
3. In Hypex, switch provider: Ctrl+Shift+P → Hypex: Select Model Provider → Claude Code CLI.

Enabling Bypass Mode

Three entry points, pick whichever feels closest:

• Ctrl+Alt+H → $(zap) Toggle Claude CLI Bypass
• Command palette: Ctrl+Shift+P → Hypex: Toggle Claude CLI Bypass Mode
• Click the $(zap) Bypass status bar badge once it appears (it only shows when claude-cli is the active provider AND bypass is on)

A modal pops up the first time you enable it explaining the trade-off. Accept and the next agent turn runs autonomously. Disable any time — the same toggle flips it back to approval mode.

Status bar badge

Whenever Bypass is ON and claude-cli is selected, a yellow $(zap) Bypass badge appears on the left of the status bar. Click to disable. The badge intentionally uses the warning background so it's never invisible — autonomous mode should always be a deliberate choice.

What gets bypassed exactly?

• Tool call approvals — read_file, write_file, edit_file, run_command, MCP tool calls all run without confirmation.
• Permission prompts — Claude Code's own permission gates (workspace boundaries, sensitive files) are skipped for this session.

What does NOT change

• Anthropic safety — Bypass is tool-level, not model-level. Claude still won't generate CSAM, malware targeting real victims, or weapons systems regardless of which interface you use.
• Workspace boundary — Hypex's resolveInsideWorkspace still rejects path traversal on tools that go through Hypex's sandbox. Bypass only affects the CLI side.
• Your subscription quotas — Bypass doesn't lift Claude Max's daily / weekly limits.

How does it compare to other autonomous agents?

Safety guidance

• Trusted workspaces only. Don't enable Bypass when reviewing a repo you don't trust — a hostile prompt or steering file could trigger destructive tool calls before you see them.
• Pair with Hypex Checkpoints. Every agent edit is a rollback point — Ctrl+Alt+Z to undo. Bypass + Checkpoints = autonomous but reversible.
• Use restrictive profiles when needed. Even with Bypass on, you can set agent.profile to default or tester for a less permissive framing.