cc-code is an open-source terminal coding agent built with Rust, claimed to be compatible with Claude Code and usable from existing .claude/ configurations with zero migration. The project address is https://github.com/cc-claws/cc-code，官方网站为 cc-claw.com.
Main capabilities
- Powered by open-source models: materials indicate it is powered by DeepSeek-V4-Pro, Mimo-2.5Pro, and GLM-5.1.
- Terminal interface: provides TUI and CLI help, supports English and Simplified Chinese. You can use /lang en or /lang zh-CN to save the language preference; when no language is saved, startup detects LC_ALL, LANGUAGE, LC_MESSAGES, and LANG, and unsupported locales fall back to English.
- Auto mode (the default permission mode, fail-closed): it does not ask the user each time, but instead makes decisions in two layers. The first layer is the deterministic layer, including user-defined rules, a built-in hard-deny list, and a read-only fast path, with zero cost and no credentials required. The second layer is the JEV semantic gate, a dedicated LLM judge (JEV_MODEL, default openrouter/typesafe/jev-1.13), accessed through an independent endpoint (JEV_ENDPOINT) and an independent key (JEV_API_KEY); it scores each call against a set of conditions, and only when a call falls into an uncertain range does it pop up a single confirmation prompt.
- Behavior without credentials: the deterministic layer still runs, with an LLM classifier as a fallback; when there is no confirmation channel (such as subagents, -p), the call is rejected. The session model does not participate in judgment; the judgment work is handled by JEV, so security checks do not consume context tokens. Shift+Tab switches between Auto and Bypass.
- Subagent cancellation (marked in the materials as unreleased): when the main agent is told to stop a certain subtask, it calls Agent(action="stop", task_id="bg-…"), canceling by exact task_id, with no fuzzy matching, avoiding mistakenly stopping similar tasks; the receipt contains a task summary, the timeline prints "• Agent "x" was stopped by main Agent", and the drawer row becomes a red solid ● and shows stopped by main Agent (keyboard stop shows stopped by user). During the wait, the main viewport shows "✻ Waiting for N background agents to finish".
- Subagent prompt card (marked in the materials as unreleased): in the main viewport, press Ctrl+O to toggle detailed mode; the startup summary of a background subtask changes from a compact tree to a classic tool card (● Agent(task preview) + └ Backgrounded agent + └ Prompt: followed by the full instruction text), and pressing Ctrl+O again collapses it.
Usage
- Installation: npm is the recommended method, with the command npm install -g @cc-claw/code; in addition, macOS / Linux script installation and Windows PowerShell installation methods are also provided.
- Binary-supported platforms: macOS (x86_64 / Apple Silicon), Linux (x86_64 / aarch64), Windows (x86_64).
- Language switching: use /lang en or /lang zh-CN.
- Permission mode switching: Shift+Tab switches between Auto and Bypass.
Applicable scenarios The directions listed in the materials include long-horizon programming, coding-driven design, swarm-based task orchestration, and end-to-end engineering work such as frontend, DevOps, performance optimization, and full-stack workflows.
Project development method The project team calls it "Nobody Coding": no human wrote a single line of cc-code code, including the architecture, the TUI, and the harness tuning that makes open-source models run reliably in the Agent loop; humans decide "what to do", and AI decides "how to do it". According to its explanation, 99% of cc-code was built this way, and recent commits come almost entirely from DeepSeek, Mimo, and GLM, with Claude appearing only in the initial stage.
License Apache License 2.0, free to use, modify, and distribute, including for commercial purposes. Version update records are in CHANGELOG; old versions from v0.6.0–v0.6.121 are also listed in CHANGELOG.
