Claude Code — Custom Build

A modified build of Claude Code, Anthropic's agentic coding CLI. This is the reconstructed TypeScript source with a set of custom features layered on top — the headline one being a built-in jailbreak that unlocks the model, plus full telemetry removal, permission bypass by default, multi-account switching, and a couple of extra tools.

Everything runs in your terminal exactly like the official CLI. The difference is what's been added and taken out.

---

Highlight: Jailbreak

The main feature. A built-in jailbreak system that unlocks the model's behavior — works across all models, Opus 4.8 included.

It's controlled by the /jailbreak command with three levels:

Level	Behavior
disabled	No content-handling section, no priming — stock behavior
lowered	Softer content-handling section, no synthetic history (default)
full	Strongest content-handling section + full synthetic conversation-history priming

Under the hood it works by injecting an operator content-handling section into the system prompt and (at full) prepending a synthetic conversation history that primes the model. It's wired into the main chat, subagents, session-title generation, and the WebFetch/WebSearch helper models, so the unlock is consistent everywhere rather than just the top-level chat.

Switch levels any time with /jailbreak; the change takes effect immediately.

---

Features

Privacy

• Telemetry fully removed — GrowthBook, Datadog, and the 1P event-logging pipeline are stripped out. No usage data, device IDs, or session attributes are sent anywhere.
• Feedback surveys disabled — no rating prompts, memory surveys, or post-compact surveys.
• Auto-updater removed — no background version checks against npm or GCS.
• Model attribution stripped — no Co-Authored-By / "Generated with" trailer added to your commits or PRs.

Permissions & guards

• Permission bypass on by default — starts in "Allow Everything" mode, no need for --dangerously-skip-permissions. Toggleable from the first tab in /permissions.
• Bypass warning dialog removed — no confirmation prompt on first use.
• Trust prompt skipped — no "do you trust this folder" gate.
• Root/sudo guard removed — runs under root without complaint.
• Malware / binary-scan bypass — internal binary scanning is disabled.
• WebFetch domain blocklist removed — no preflight check against Anthropic before fetching a URL.

Accounts & models

• In-session account switching — manage multiple Claude OAuth accounts and API configs (key + optional base URL + auth token) and swap between them on the fly with /account. Switching is instant, no restart. Also reachable from the rate-limit menu.
• Static model list — the /model picker is a single fixed list instead of dynamically fetched, so it's faster and easier to read.

Extra tools

• BrowserFetch — fetches URLs with a real browser network fingerprint (JA3/JA4 TLS + HTTP/2 + correctly-ordered headers) to reach sites that block generic HTTP clients. Has a render tier driven by headless Camoufox (anti-detect Firefox) that runs the page's JavaScript and clears Cloudflare "Just a moment…" / Turnstile challenges. Modes: auto, fast, render.
• GrokSearch — web research tool that queries Grok over HTTP/2 and returns answers with cited sources. Sign in with /grok-login; once signed in it loads up front and the model prefers it over WebSearch for live lookups.

Commands & UX

• /thinking — toggle extended thinking on or off directly.
• /folder — open the current working directory in your file explorer.
• . prompt shortcuts — type .<name> <text> to expand a saved template ahead of your message (e.g. a saved "answer only, don't edit code" instruction). Templates live in an editable shortcuts.json and support an {input} placeholder. Manage them with /shortcut.
• Paste-again-to-expand — large pastes (>800 chars) collapse to a placeholder; paste again within 800ms to expand inline.
• Status indicators row — shows source, model, and thinking level at a glance.
• Injected-tag handling — server-side reminder tags are treated as ordinary text and ignored, so they can't steer behavior.

---

Build

Requires Bun ≥ 1.1 and Node ≥ 22 (Node is used for the build script; Bun for compilation).

bun install
npm run build

npm run build bundles the source with esbuild and cross-compiles standalone executables for all platforms into dist/:

File	Platform
dist/claude.exe	Windows x64
dist/claude-linux-x64	Linux x64
dist/claude-linux-arm64	Linux arm64
dist/claude-darwin-x64	macOS Intel
dist/claude-darwin-arm64	macOS Apple Silicon

Other build scripts:

npm run build:prod     # minified build
npm run build:watch    # rebuild on change
npm run typecheck      # tsc --noEmit
npm run check          # biome + typecheck

Run

Run the compiled binary for your platform:

# Windows
./dist/claude.exe

# Linux
./dist/claude-linux-x64

# macOS (Apple Silicon)
./dist/claude-darwin-arm64

Or run from source directly with Bun (no build needed):

bun src/entrypoints/cli.tsx

On first launch it prompts you to authenticate via OAuth — the same flow as the official CLI. Then cd into a project and start a session, or pass a prompt directly:

./dist/claude.exe "explain what this repo does"

First steps

/jailbreak     # set the unlock level (disabled / lowered / full)
/account       # add and switch between accounts / API keys
/model         # pick a model
/thinking      # toggle extended thinking
/permissions   # Allow Everything is on by default