# Myeongpil (명필) — Unity editor task delegation for AI agents

> Your agent plans; Myeongpil executes the repetitive Unity-editor steps at a fraction of the
> per-step token cost. It is an operator model, not a chat model: given a goal and the steps taken
> so far, it returns the SINGLE next Unity tool call as JSON.

Auth: X-API-Key header. Free key (0.5M tokens/mo): POST /signup {"email": "you@example.com"}
Console (usage, billing, setup): https://myeongpil.14dimension.com/console
Machine-readable spec: /openapi.json
<!-- Codex, 2026-08-04: 최초 PyPI 공개 이후 에이전트가 긴 서버 wheel URL을 만들지 않도록 정식 설치 경로를 명시한다. -->
PyPI package: https://pypi.org/project/myeongpil-mcp/ · Run: uvx myeongpil-mcp

## How to use it (two ways)

1) **MCP (recommended)** — register the local shim so one delegation runs many steps without
   your own tokens paying for each round trip. Setup commands with your key are in the console.
   Works today with Claude Code, Claude Desktop, and Codex CLI (they support local MCP servers).
   Requires MCP for Unity (https://github.com/CoplayDev/unity-mcp) connected to your editor.

2) **Direct API** — call /v1/next-step yourself and execute the returned tool call.
   You keep the loop; useful if you already drive Unity and only want the decision.

## Core endpoints
- POST /v1/next-step {"goal": "...", "history": [{"tool": "...", "input": {...},
  "result": "<what came back>", "is_error": false}]}
  → {"tool": "execute_code", "input": {...}, "usage": {...}}
  Always send `result` for each history step — without it the model cannot tell whether the last
  action succeeded, and picks the wrong next step. `result_head`/`output` are accepted as aliases.
- POST /v1/warmup — wake the GPU worker ahead of time (serverless; first call after idle is slower).
- GET /v1/me — this month's token usage and remaining quota.
- GET /billing/status · POST /billing/checkout · GET /billing/history — plans, top-ups, invoices.

## Limits and behaviour agents should know
- The model returns ONE tool call per request. You execute it and send the result back.
- It is trained on real Unity editor trajectories, so it continues an established pattern well.
  It is weaker at deciding *when to switch* activity (e.g. run code vs. check the console) —
  keep that judgement on your side and hand Myeongpil well-scoped, repetitive work.
- 402 means quota exhausted or the user's monthly spending cap was reached — surface it, don't retry.
- 503 BACKEND_WARMING means the GPU is starting; retry in ~2 minutes. Tokens are not charged.

## Sibling products (same key system, same API shape)
- Images / UI sprites / icons — 혜안 HyeAn: https://hyean.14dimension.com
- Sound effects / audio — 청명 Cheongmyeong: https://cheongmyeong.14dimension.com
- Particle VFX (Unity prefabs) — 휘영 Hwiyeong: https://hwiyeong.14dimension.com
- Korean document extraction — 속기 Sokgi: https://sokgi.14dimension.com
- Document search — 온고 Ongo: https://ongo.14dimension.com
- Voice-recording search — 경청 Gyeongcheong: https://gyeongcheong.14dimension.com
- AI voices (unmetered TTS, batch game dialogue) — 정어 Jeongeo: https://jeongeo.14dimension.com
Keys are per-product (a 혜안 key will not work here). Each has its own /llms.txt and free tier.

Brand: https://ai.14dimension.com · Contact: admin@14dimension.com