Give Wukong a goal. It loops until the change is ready.

Wukong Code is a terminal AI coding agent built around one workflow: Goal → Write → Check → Review → Fix.

Install v0.1.1 Download Wukong Code View the release

---

One workflow

Loop until ready — /loop works on one goal, runs the repository’s real checks, reviews from a fresh read-only context, and fixes blocking findings.

Stop instead of drift — v0.1.1 keeps the goal fixed, remembers earlier blockers, and stops with a clear reason when another iteration would repeat the same work.

Resume unfinished work — continue local Codex, Claude Code, Cursor, Kimi Code, or Grok sessions without modifying the source session or replaying old tools.

Experimental roles and subagents — after explicitly enabling experimental.role_profiles, transform Wukong for focused work, delegate an independent task, and inspect background agents without creating a second workflow. They are not part of the default command surface.

BYOK — choose your provider and model API. Wukong does not require a local model. Device Login is optional and does not gate local Loops.

Install

macOS and Linux:

curl -fsSL https://wukong.today/install.sh | sh
wukong --version

Windows users can download the matching x64 or ARM64 ZIP from the release page. Download the adjacent .sha256 file too, then follow the exact verification, extraction, and PATH steps in Getting Started.

Start a Loop

# Start the TUI
wukong

# Or preview a headless Loop; then add its exact Headless start flags
wukong loop "finish the current change" --dry-run

Inside the TUI:

/provider
/loop add input validation to the signup form
/resume codex
/resume kimi

Every completed Gate finishes as PASS, NEEDS_WORK, or ERROR. TUI Ctrl-C or /loop pause is resumable PAUSED; /loop stop or a headless interruption before a Gate verdict is reported separately as STOPPED_BY_USER. Start with the Getting Started guide or browse the Command Reference.

Current product boundary

The v0.1.1 release keeps local Loops free and bring-your-own-key, without a Guest trial, sign-in requirement, or monthly account limit. The default per-run safety limit is 10 iterations and can be changed explicitly. There is no public paid plan, Checkout, hosted report workflow, hosted inference, or managed model credit.