dualface/kander

English | 简体中文 | 日本語

Strict rule-driven multi-agent parallel development with built-in independent review and delivery gates to ensure automated delivery quality.

Born from real engineering
Since August 2026, Kander has driven more than 1,700 real tasks across 5 projects, 1,317 of them in the QuickTUI production environment. Refined through handling actual code conflicts, race conditions, and complex bugs, it establishes truly reliable agent orchestration, independent review, and crash recovery. Even with lower-cost models, it guarantees delivery quality.

One developer + 6 agents	68 days	5 projects
1,726 task cards done	1,694 independent review runs	49% of review batches found must-fix issues

Numbers and definitions: Production Stats (as of 2026-10-07)

In-depth reading: Kander in Production | Full Production Retrospective (Deep Dive)

• Two-Stage Independent Review Gates: Physical isolation between execution and review. PMQA catches semantic and state-machine bugs, blocking false-green tests.
• Conflict-Aware Scheduling: Static analysis of Git branch overlaps and file modification boundaries ensures controlled concurrency without write conflicts.
• Terminal and Workspace Isolation: Isolated parallel agent execution powered by Git Worktrees and tmux/herdr containers.
• Zero-Token GitHub Integration: Seamlessly reuses local gh credentials to browse issues, import task cards, and sync progress directly from the board.

Running requires Git, plus at least one of Codex, Claude, Grok, Cursor, or Pi.

macOS — install with Homebrew:

brew install dualface/tap/kander

kander

Linux — download the binary directly:

ARCH=$(uname -m); [ "$ARCH" = x86_64 ] && ARCH=amd64; [ "$ARCH" = aarch64 ] && ARCH=arm64
curl -fsSL "https://github.com/dualface/kander/releases/latest/download/kander-linux-${ARCH}.tar.gz" | tar xz

./kander

Windows — download kander-windows-amd64.zip from Releases, unzip it, and run kander.exe.

On first launch, if not yet installed, an interactive wizard starts. Once installation finishes, it is ready to use.

Four steps to get going:

1. Start an agent session and discuss the requirement or task there, making the goal and acceptance criteria clear. The agent's Plan mode is recommended.
2. Once the task is confirmed, the agent (with Kander rules loaded) confirms whether to launch it through the kanban flow. Confirm, and the task is created and started automatically.
3. When you have multiple requirements, repeat steps 1-2 for each one, continuously scheduling and launching tasks.
4. Check task status with the command-line interface:

kander

The board contents above come from my real project https://quicktui.ai. QuickTUI is a tool for remotely operating the agents on your computer; it supports iOS/Android/macOS/Linux/Windows and is free to use.

Further reading: How to Advance Tasks Efficiently (PDF slides).

• kander: Open the interactive terminal kanban board.
• kander doctor: Check and repair environment dependencies, agent availability, launchers, and rules configuration.

Linking a project to a GitHub repository needs the GitHub CLI (gh). Kander never asks for, reads, or stores a token; it reuses the credentials gh already manages.

On the terminal board, press g to open the GitHub issue list in an overlay.

A: Stop the agent currently working on the card, launch a new agent, and prompt it to take over and continue task card TASK-ID (replace TASK-ID with the actual card ID). In the terminal kanban board, select the card and press y to copy its task ID.

A: Copy the task card ID, launch an agent, and instruct it to act as the coordinator to take over and advance the task group containing TASK-ID.

A: Start an agent (with Kander rules loaded) and directly ask for the current status and progress of any unfinished task cards.

If you read agent reports in Chinese, pair Kander with ste-zh. It is an agent skill that makes the agent report results by ASD-STE100 (Simplified Technical English) principles, in Chinese.

In the author's daily Kander workflow, this combination works very well. Kander's completion reports already require actual verification results; ste-zh makes every reply easy to scan:

• the conclusion comes first;
• status words are fixed, such as "completed", "not verified", and "blocked";
• every conclusion states whether it was verified, and how;
• when you must decide, the options are numbered.

When many cards run in parallel, you can read each report in seconds and know what is done, what is unverified, and what needs your decision.

Install it for Claude Code (the directory name must be ste), then type /ste in a session, or load it from your global agent rules so every session uses it:

git clone https://github.com/dualface/ste-zh.git ~/.claude/skills/ste

• Review Disposition & Completion Gate
• Card Transactions & Crash Recovery
• Terminal Backends & Container Definitions
• Durable Dispatch Protocol
• GitHub Issue Import & Result Protocol

This project is under the MIT License; see LICENSE.

Release notes live in CHANGELOG.md.

Other projects by dualface, the author of Kander:

• ste-zh: an agent skill that makes agents report results in Chinese by ASD-STE100 principles: conclusion first, fixed status words, explicit verification state.
• Ullage: a local daemon and CLI that shows subscription usage for Claude, ChatGPT, Grok, Cursor, and more.
• QuickTUI: a full terminal for any coding agent, on your phone. Self-hosted, direct connection. Free for a single host.