cc-to-grok-bridge

English | 繁體中文 | 简体中文

Bridge Claude Code → Grok Build: reuse rules / skills, run the same hook hard-gates via an adapter, optional memory mirror. Grok-only assets never enter Claude’s context.

What this is

Research helper: ccgrok.sh

A small wrapper that drives grok -p as a disciplined research tool. It prepends a model-agnostic research-method prefix (research-prefix.md) and runs from a clean temp directory, so a pure research query is not polluted by project rules.

scripts/ccgrok.sh "your question"

The prefix forces sourced, dated answers with facts separated from speculation, and (measured on the same query) makes Grok flag stale or downgraded facts instead of stating them with false confidence. The same research-prefix.md works with any headless AI CLI (grok -p, agy -p, claude -p).

The read-only flags (--disallowed-tools "run_terminal_cmd,write,search_replace") keep the run unable to touch local files even under auto-approve.

Compatibility matrix

In practice: moving a Claude Code setup onto Grok Build is usually much smoother than the Antigravity / Gemini bridge path (real hard-blocks + simpler memory). Always-on plugins still need a deliberate path.

Plugins note (Low = not drop-in, but documented fixes exist):
CC auto-enabled plugins ≠ Grok per-session injection. A: grok plugin install only when upstream has Grok packaging (SessionStart always-on). B: otherwise put always-on text in .grok/rules/ (auto-load, no slash). See docs/06_plugins.md.
Outlook: if most common plugins ship on the Grok marketplace with real Grok packaging, compatibility can rise to medium–high (still not auto-porting CC settings, but path A becomes normal and B is the exception).

Memory detail: bridge day-use does not require [memory] enabled=true first — the rules pointer loads with the project. Product memory improves search if you want it.

MCP detail: ask an AI coding agent to follow AGENTS.md (Notion, Google, OAuth). You approve the browser; you should not be asked to paste long terminal homework.

What the hook adapter does (plain language)

Claude Code hooks are shell scripts (e.g. block reading .env). Grok also runs hooks, but the contracts differ:

1. JSON field names on stdin differ (payload shape)
2. How “block this tool call” is reported differs (deny protocol)

So you cannot point Grok straight at CC scripts. A thin adapter (scripts/hook_adapter.py) sits in the middle: Grok calls it; it calls your existing ~/.claude/hooks/*.sh.

Why this design:

• Single source of truth — blocklists stay in CC hooks
• Hard block on Grok — not “model please remember”; the tool call is actually denied
• No forking CC scripts — no need to edit ~/.claude/hooks/*.sh just for Grok

Requirements

• macOS or Linux with Python 3.10+
• Claude Code already set up (hooks under ~/.claude/hooks/, skills under ~/.claude/commands/)
• Grok Build CLI / TUI
• No API key required for the bridge scripts themselves

Quick start

git clone https://github.com/<YOUR_USERNAME>/cc-to-grok-bridge.git
cd cc-to-grok-bridge

# From your real workspace root (the folder that has CLAUDE.md / .grok/)
export CC_GROK_WORKSPACE=~/path/to/your-workspace

python3 scripts/install_bridge.py
python3 scripts/memory_sync.py --workspace "$CC_GROK_WORKSPACE"
python3 scripts/bridge_doctor.py --workspace "$CC_GROK_WORKSPACE"   # expect fails=0

Then open Grok from that workspace and restart the session so hooks reload.

Layout

Security

• Scripts never copy MCP env / tokens / OAuth secrets.
• Sensitive-read hard block is verified via adapter + your CC block_sensitive_read hook (e.g. .clasprc*, .env).
• Grok-only rules live under <workspace>/.grok/rules/ and must not be written into Claude’s system files by this bridge.

Limitations

1. MCP: secrets never auto-copied; claude.ai cloud connectors not portable — reinstall via docs/03_mcp.md / AGENTS.md
2. Global Grok ~/.grok/memory/MEMORY.md (/remember) ≠ CC project memory; sync uses a project subfolder
3. After install, restart the Grok session from the workspace root

License

MIT — see LICENSE.