ClawGod

English | 中文 | 日本語

God mode for Claude Code.

This is NOT a third-party Claude Code client. ClawGod is a runtime patch applied on top of the official Claude Code. It works with any version — as Claude Code updates, ClawGod automatically re-extracts and re-patches against the new version on the next launch.

Prerequisites

Install these before running the ClawGod installer:

Tool	Why	Install
Claude Code (native binary)	ClawGod patches the official Bun standalone binary you already have	claude.ai/install.sh (macOS/Linux) or claude.ai/install.ps1 (Windows)
ripgrep	Required by Claude Code's Grep tool	brew install ripgrep / apt install ripgrep / winget install BurntSushi.ripgrep.MSVC
Node.js >= 18	Used by the patcher	nodejs.org
Bun	Runtime for the patched cli.js; auto-installed if missing	bun.sh, npm install -g bun, scoop install bun, or choco install bun

Install

macOS / Linux:

curl -fsSL https://github.com/0Chencc/clawgod/releases/latest/download/install.sh | bash

Windows (PowerShell):

irm https://github.com/0Chencc/clawgod/releases/latest/download/install.ps1 | iex

Green logo = patched. Orange logo = original.

What it does

Feature Unlocks

Patch	What you get
Internal User Mode	24+ hidden commands (/share, /teleport, /issue, /bughunter...), debug logging, API request dumps
GrowthBook Overrides	Override any feature flag via config file
Agent Teams	Multi-agent swarm collaboration, no flags needed
Computer Use	Screen control without Max/Pro subscription (macOS)
Auto-mode	Unlocks auto-mode for third-party API users (no firstParty gate)
Classifier tuning	Tunable auto-mode classifier: timeout / model / retries (CLAWGOD_CLASSIFIER_TIMEOUT_MS, CLAWGOD_CLASSIFIER_MODEL, CLAWGOD_CLASSIFIER_RETRIES)
Ultraplan	Multi-agent planning via Claude Code Remote
Ultrareview	Automated bug hunting via Claude Code Remote

Restriction Removals

Patch	What's removed
CYBER_RISK_INSTRUCTION	Security testing refusal (pentesting, C2, exploits)
URL Restriction	"NEVER generate or guess URLs" instruction
Cautious Actions	Forced confirmation before destructive operations
Login Notice	"Not logged in" startup reminder

Geo-Steganography Neutralization

Patch	What's neutralized
Date String (qla)	System prompt encodes user location via Unicode apostrophe variants (U+0027 / U+2019 / U+02BC / U+02B9) and date separator (- vs / for CN timezone). Patched to always use ASCII ' and unmodified date format
Geo-Detection Probe (rdp)	Client-side three-axis detection: timezone (Asia/Shanghai / Asia/Urumqi), proxy hostname against XOR-obfuscated 100+ domain blocklist, CN-LLM vendor keywords in base URL. Patched to always return null
Apostrophe Selector (odp)	Selects one of four Unicode apostrophes based on detection results. Patched to always return ASCII ' (defense-in-depth)

Visual

Patch	Effect
Green Theme	Brand color → green. Patched at a glance
Message Filters	Shows content hidden from non-Anthropic users

Reliability

Feature	What it does
Glob/Grep Restore	Bun compile inlines EMBEDDED_SEARCH_TOOLS=true, hiding built-in Glob/Grep tools. Patch un-inlines the env check and adds bfs/ugrep binary availability detection — tools are restored when running under Bun runtime
Renderer Shim (2.1.271+)	Claude Code 2.1.271 builds its TUI renderer on Bun.ant.CellSegmenter, a private API that only Anthropic's bundled Bun exposes. ClawGod loads a JS implementation before the patched bundle, so the TUI paints under stock Bun instead of stalling before the first frame
1h Prompt Cache	Forces 1h TTL allowlist on (was effectively 5m → much higher cache_creation token usage)
Third-Party Cache Fix	Auto-disables x-anthropic-billing-header when baseURL is non-Anthropic. The header's per-request cch field breaks prompt-cache hit rate on DeepSeek / OneAPI / Bedrock / vLLM and any other Anthropic-compatible proxy. You no longer need to set CLAUDE_CODE_ATTRIBUTION_HEADER=0 yourself.
Auto Re-patch	Detects when the user's native Claude binary has been upgraded; transparently re-extracts and re-patches on next launch
Update Notification	Checks GitHub releases once per 24h (async, non-blocking). Shows a one-line notice if a newer ClawGod version is available
Lean Settings	Three-level token optimization for ~/.claude/settings.json. on (default): removes unused tool definitions + disables Workflows/Artifact; Remote Control remains available. max: additionally disables Remote Control and removes Plan mode, Agent Teams, bundled skills. off: all tools restored

Lean Settings are non-destructive and persist across updates. Toggle anytime: claude --lean-on (default) / claude --lean-max (aggressive) / claude --lean-off (restore all). To opt out of a single setting, set it yourself (e.g. "disableArtifact": false).

Remote Control (/remote-control, /rc) is allowed in on/off and disabled by default only in max. Updating in on mode or running claude --lean-on clears the older Lean Remote Control disable setting. Only max defaults CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1; explicit environment restrictions are preserved. Remote Control still requires upstream account, authentication, endpoint, and organization eligibility.

Commands

claude              # Patched Claude Code (replaces the official launcher)
clawgod             # Same as `claude`, explicit & guaranteed entry point
claude.orig         # Original unpatched version (auto-backed-up)

clawgod is unambiguous: on Windows where claude.exe may shadow claude.cmd, clawgod.cmd always works. Even after official self-update overwrites claude, clawgod keeps running the patched build.

Configuration

~/.clawgod/provider.json is auto-created on first run. Setting apiKey lets you skip OAuth entirely and point ClawGod at any Anthropic-compatible endpoint.

{
  "apiKey": "sk-ant-...",
  "baseURL": "https://api.anthropic.com",
  "model": "",
  "smallModel": "",
  "effort": "",
  "timeoutMs": 3000000
}

• apiKey set → ClawGod injects it as ANTHROPIC_API_KEY and isolates from ~/.claude/settings.json. The default protocol requires an Anthropic Messages-compatible endpoint. For Chat Completions gateways, use the configuration below. A non-Anthropic baseURL populates only ANTHROPIC_AUTH_TOKEN for gateway auth.
• apiKey empty → OAuth path. Run claude auth login once; ~/.claude keeps hosting your subagents, skills, and MCP settings.
• effort → Sets reasoning effort; an existing CLAUDE_CODE_EFFORT_LEVEL takes precedence. With protocol: "openai-chat", type: "grok" or "openai-compat", the proxy sends reasoning_effort even when Claude omits effort for a custom model alias. low, medium, high, and xhigh pass through; max maps to xhigh; auto omits the parameter to use the upstream default. Choose a level supported by your upstream model. Empty/unset configuration leaves request-level effort in control and adds no effort parameter when the request has none.

OpenAI Chat Completions endpoints

For an endpoint that exposes /v1/chat/completions, configure:

{
  "protocol": "openai-chat",
  "apiKey": "sk-...",
  "baseURL": "https://example.com/v1",
  "model": "your-upstream-model",
  "smallModel": "your-upstream-model"
}

baseURL is the API base (including /v1 when required), not the full /chat/completions URL. Set the model names to IDs accepted by your provider. ClawGod starts a loopback proxy inside the launcher process; no external gateway or separate service is required. timeoutMs also covers the upstream request and stream; an existing API_TIMEOUT_MS takes precedence.

• Omitting protocol keeps the existing behavior. type: "openai-compat" and type: "grok" remain supported; an explicit protocol takes precedence. protocol: "anthropic" uses Messages directly. Grok defaults to https://api.x.ai/v1 and retains its settings/environment API key fallback. Other Chat providers require an explicit API base and key.
• Supports text, system prompts, tool calls/results, forced tool selection and parallel-tool controls, streaming, usage, and base64/URL images. Parallel tool arguments are assembled and validated before their content blocks are emitted; text still streams immediately. Missing upstream usage remains zero rather than a fabricated exact count.
• Base64 PDFs are sent as Chat Completions file parts and require file support in the upstream model/API. Text documents are sent as text. PDF URLs, document citations, non-text tool results, server tools, and structured output formats are rejected with a clear error. Historical Anthropic thinking/signature blocks are omitted. Model-specific reasoning features are not losslessly translated.
• /v1/messages/count_tokens estimates locally without a billed generation request. The x-clawgod-token-count: estimate response header marks the result. It uses UTF-8 text bytes / 3 plus message overhead, 1600 tokens per image, and decoded PDF bytes / 3. This is a rough budget, not the model's tokenizer; PDF byte size does not reflect page count and the estimate cannot guarantee context-window fit.
• Chat stop maps to end_turn (or tool_use when the response contains tool calls): the protocol does not distinguish natural completion from a matched stop sequence. length, tool_calls, and content_filter map to max_tokens, tool_use, and refusal. Upstream HTTP errors retain their status and Retry-After; malformed or truncated streams produce an error.
• /v1/responses and automatic protocol detection are not supported. Use openai-chat only for Chat Completions endpoints.

Feature Toggles

~/.clawgod/patches.json (auto-created empty) switches features off persistently — your choices survive updates and reinstalls. Absent key = on.

{ "theme": false, "geo-neutralize": false }

Feature id	Controls
agent-teams	Agent Teams always enabled
computer-use	Computer Use unlock
ultraplan	Ultraplan slash command
ultrareview	Ultrareview slash command
voice-mode	Voice Mode
auto-mode	Auto-mode model selection on third-party APIs
classifier-tuning	Auto-mode classifier overrides: CLAWGOD_CLASSIFIER_TIMEOUT_MS (min deadline; unset = 60-120s by context, v2.1.251+), CLAWGOD_CLASSIFIER_MODEL, CLAWGOD_CLASSIFIER_RETRIES (unset = 4)
theme	Green brand/logo color scheme
geo-neutralize	Geo/proxy steganography neutralization in system prompt
cyber-risk	Removes CYBER_RISK_INSTRUCTION from system prompt
url-restriction	Removes URL generation restriction from system prompt
cautious-actions	Removes "Executing actions with care" section from system prompt
not-logged-in	Removes "Not logged in" notice
message-filter	Bypasses non-ant message/attachment filters
bun-ant-shim	Bun.ant.CellSegmenter renderer shim for Claude Code 2.1.271+ (see Reliability)

For a single launch, set an env var instead — feature id upper-cased, dashes to underscores:

CLAWGOD_FEATURE_THEME=false claude     # green theme off, this run only
CLAWGOD_FEATURE_GEO_NEUTRALIZE=true claude  # temporarily re-enable one disabled in patches.json

How it works

Since @anthropic-ai/claude-code v2.1.113, the npm package no longer ships cli.js — it's a thin loader that dispatches to platform-specific Bun standalone binaries. ClawGod adapts:

1. Locates the user's installed native Bun binary in ~/.local/share/claude/versions/
2. Extracts the embedded cli.js source from the __BUN segment (Mach-O / ELF / PE)
3. Extracts the embedded .node native modules (audio-capture, image-processor, computer-use-*, url-handler) into ~/.clawgod/vendor/
4. Rewrites /$bunfs/... virtual paths to point at the extracted modules
5. Applies 29 regex-based patches (version-agnostic — same patches work across many releases)
6. The claude / clawgod launchers run the patched cli.js under the Bun runtime
7. Loads bun-ant-shim.cjs before the patched cli.js, supplying the Bun.ant.CellSegmenter renderer API that Claude Code 2.1.271+ expects

A .source-version stamp in ~/.clawgod/ records which native version was patched. On every launch the wrapper compares it against the latest binary in versions/; if the user upgraded Claude Code via the official installer, ClawGod auto-re-patches on the next run.

Update

Just run claude update as usual. ClawGod patches the command to route through its own installer, which pulls the current Anthropic release from npm (@anthropic-ai/claude-code-<plat>@latest), re-extracts cli.js, re-applies patches, and rewrites the launcher. So the upstream update command keeps working the way you expect — you get the latest Claude, with patches still applied, in one step.

Extra options:

claude update --version 2.1.180   # Pin to a specific Claude Code version
claude update --no-upgrade        # Re-patch the installed Claude version

--version is useful when a new release has issues and you want to stay on a known-good version. --no-upgrade re-applies the latest patches to a complete clean-source backup of the installed version. Older installations without that backup download the same Claude version once to recover it; subsequent re-patches reuse the local backup.

If you'd rather invoke the installer directly (same effect, both paths fetch the same upstream release and re-patch):

macOS / Linux:

curl -fsSL https://github.com/0Chencc/clawgod/releases/latest/download/install.sh | bash

Windows:

irm https://github.com/0Chencc/clawgod/releases/latest/download/install.ps1 | iex

If you'd rather drop ClawGod and use Anthropic's original claude update (which manages its own paths and would overwrite our launcher), uninstall first:

bash ~/.clawgod/install.sh --uninstall

Uninstall

macOS / Linux:

curl -fsSL https://github.com/0Chencc/clawgod/releases/latest/download/install.sh | bash -s -- --uninstall
hash -r  # refresh shell cache

Windows:

irm https://github.com/0Chencc/clawgod/releases/latest/download/install.ps1 -OutFile install.ps1; .\install.ps1 -Uninstall

Uninstall restores claude.orig → claude and removes the clawgod alias.

After install or uninstall, restart your terminal or run hash -r if the command doesn't take effect immediately.

License

GPL-3.0 — Not affiliated with Anthropic. Use at your own risk.

Star History