ZwiCode

A Hermes-style Castle CLI game development agent. One setup script installs a Python agent at ~/.zwicode/ that talks to LLM providers and drives castle-cli to build, serve, and iterate on Castle deck games.

Quick start

Works on Linux, macOS, and Android (Termux). Requires Node 18+ and bash (on Termux: pkg install nodejs first).

npx @zwiq/zwicode

First run executes the setup script (installs deps, writes the agent, bundles Castle reference docs + design guides, creates the zwicode launcher), then drops you into the agent.

Or install globally:

npm install -g @zwiq/zwicode
zwicode

zwicode-setup re-runs just the setup/upgrade step.

What's new in 1.3

• Design guides — six bundled design docs (design/2d-games, design/3d-games, design/ui, design/game-ui-ux, design/character-design, design/animation-design) that the agent is instructed to read before writing game/UI/character/animation code, with pre-build checklists. Output quality jumps from "placeholder" to designed.
• Multi-blueprint freedom — the system prompt no longer pushes a single actor/blueprint first pass; the agent creates as many blueprints (player, enemies, projectiles, HUD, controller...) as the design needs from the start.
• Per-model temperature — vendor-recommended sampling per model (e.g. GLM 5.2 → 1.0, Kimi K2.6 → 0.6, Qwen 3.8 2.4T → 0.7) instead of one global value, fixing temperature-induced loops and hard failures.
• Anti-loop chunked writes — write_file_chunk echoes back the last lines written so the model always has ground truth.
• No silent doc truncation — docs are served up to 24 KB with an explicit truncation notice instead of being silently cut at 8 KB.
• DeepSeek provider removed.

Providers

| Provider | Models | |--------------|-------------------------------------------------------| | Zwiq Portal | qwen-3.8-2.4t-a95b (default), qwen-3.8-27b, minimax-m3, glm-5.2 | | Anthropic | claude-sonnet-5, claude-opus-4-8, claude-haiku-4-5 | | NVIDIA | nemotron-3-ultra-550b, kimi-k2.6, glm-5.2 |

Set keys via environment variables or ~/.zwicode/.env (ZWIQ_PORTAL_API_KEY, ANTHROPIC_API_KEY, NVIDIA_API_KEY). You can also export ZWIQ_PORTAL_API_KEY before running setup to have it written for you.

The flagship qwen-3.8-2.4t-a95b runs with 131,000 max output tokens and a 262,000-token extended-thinking budget.

Features

• Hermes-style terminal UI: gradient wordmark, live status bar with real context/output/thinking token counts, bordered tool-result panels, ❯ prompt
• Castle reference docs and design guides bundled and served to the model on demand
• Tools: run_bash (with approval), read_file, chunked write_file_chunk, delete_file, plus user-defined custom tools and skills
• Streaming, session history, auto-approve toggles, custom OpenAI-compatible providers via /provider
• Slash commands: /clear /model /provider /settings /history /tools /skills /docs /help /quit

Requirements

• Linux or macOS (Termux on Android also works)
• python3, nodejs (for castle-cli), jq, curl

License

MIT