convez.ai
AgenteAutomaçãoAvançado
umyunsang

UMMAYA – Agent Multi-Ministério para Administração

umyunsang/UMMAYA

Terminal agent que traduz linguagem natural em ações de serviços públicos coreanos

Ver no GitHub
BomPluginScore · de 0 a 100

Agent CLI conversacional open-source que simplifica workflows de serviços públicos coreanos usando linguagem natural. Funciona como intermediário inteligente entre cidadãos e infraestrutura governamental, interpretando requisições em coreano e roteando através de ferramentas especializadas (busca, localização, verificação e envio). Integra dados de múltiplas agências: JUSO, SGIS, data.go.kr, registros de hospitais e canais de emergência.

Instalação

$npx convez-ai umyunsang/UMMAYA

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
5
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 1 mês
Versão
v0.2.30
Issues abertas
31
Contribuidores
3

Casos de uso

  • Encontrar emergência médica próxima com resolução automática de localização
  • Navegar procedimentos administrativos complexos de múltiplas agências
  • Consultar dados públicos sem saber qual portal acessar

Veredito

Projeto técnico bem-executado com real valor para usuários coreanos, mas nicho extremamente específico (linguagem, contexto governamental, APIs regionais). Excelente referência arquitetural para agents multi-ferramenta, porém com aplicabilidade limitada fora do ecossistema administrativo coreano.

README

A tradução para o português deste README ainda está na fila.

Click the demo to open the MP4. The recording comes from a real ummaya terminal session captured with t-rec; waiting time was trimmed, but prompts, UI, tool calls, and answers were not synthesized.

UMMAYA

npm version license docs

Unified Multi-Ministry Agent for Your Administration. 읽으면, 엄마야.

UMMAYA is a terminal agent for Korean public-service workflows. You describe the outcome in natural language, and UMMAYA routes the request through four main tools: find, locate, check, and send.

Read the full documentation at ummaya-docs.pages.dev/ko. Start with the Korean user guide if you want the product purpose first, or use llms.txt when an agent needs machine-readable project context.

npm install -g ummaya

Homebrew on macOS:

brew install --cask umyunsang/ummaya/ummaya

The Homebrew cask installs a prebuilt macOS archive and does not run npm install or bun install during installation. The unqualified official-cask form will work only after Homebrew accepts UMMAYA into Homebrew/homebrew-cask.

Run ummaya, type /login, and paste your FriendliAI API key. Public CLI users should not need to prepare Kakao, JUSO, SGIS, data.go.kr, or other public API keys.

UMMAYA is an academic and R&D project. It is not an official service of Anthropic, LG AI Research, FriendliAI, the Korean government, or any public agency.

First Run

ummaya

Inside the terminal session:

/login

Then ask in everyday language:

동아대 승학캠퍼스에서 친구가 갑자기 아프면 지금 바로 연락할 응급실 어디가 가까워?

UMMAYA stores sessions locally. When a session can be resumed, the terminal prints a command such as ummaya --resume <session-id>.

Why UMMAYA

Korean public services already exist across government portals, public datasets, address systems, emergency channels, weather services, hospital registries, and identity rails. The hard part for a citizen is not whether those systems exist. The hard part is knowing which one to use, in what order, and what to do next.

People do not usually think like this:

Call the emergency medical institution API after resolving my administrative district code.

They think like this:

다대1동 근처에서 지금 전화해볼 수 있는 병원 어디야?

UMMAYA treats that sentence as the interface. The terminal should show the work as it happens: location resolution, public-service lookup, protected-work checks when needed, compact tool results, and a final answer that a person can act on.

What You Can Ask

These are prompts you can type directly into UMMAYA. They are written as user tasks, not as API names.

SituationPrompt
Nearby emergency care동아대 승학캠퍼스에서 친구가 갑자기 아프면 지금 바로 연락할 응급실 어디가 가까워? 찾아진 곳만 이름, 주소, 전화번호로 정리해줘.
Local clinic lookup다대1동 근처에서 오늘 전화해볼 수 있는 내과가 있을까? 가까운 곳 위주로 주소랑 전화번호를 알려줘.
Weather for a walk퇴근하고 다대포해수욕장 걸어가도 괜찮을까? 지금 비 오는지랑 체감상 추운지만 알려줘.
Weather plus road risk비 오는 날 다대포에서 김해공항까지 차로 가야 해. 날씨랑 도로 위험 정보를 같이 보고 조심할 점을 알려줘.
Moving workflow이사했어. 전입신고하고 자동차, 건강보험, 학교 관련 주소 변경까지 가능한 단계는 진행하고, 공식 권한이 필요한 단계는 어디서 멈추는지 알려줘.
Permissioned workflow모바일 신분증으로 본인확인하고, 내 자격으로 가능한 복지 지원을 찾아서 신청 가능한 단계까지 진행해줘.

Live public-information flows are shown most often in the demo because they can be verified through public APIs. Identity, submission, payment, and correction flows are permissioned; UMMAYA routes them through check first, receives a scoped delegation context when the identity channel allows it, and only then lets send call the wrapped public-service channel. Today the public identity-check tools are mock or shape-mirrored unless an official credentialed channel is available; the OpenDID hackathon work targets the conversion of that check path into a live OmniOne CX / Open DID / Chain-backed tool. When official live authority is not available, the same flow is marked Mock or Handoff instead of pretending that a protected government action was completed.

Main Tools

UMMAYA keeps the model-facing surface small. Instead of exposing dozens of agency-specific APIs directly to the LLM, it gives the agent four durable primitives and lets adapters handle ministry-specific details behind them.

locate

locate turns places, addresses, coordinates, and administrative areas into usable location context. A prompt like "동아대 승학캠퍼스 근처" may need coordinates, district names, and region codes before any public-service lookup can make sense.

Typical use:

동아대학교 승학캠퍼스 근처에서 지금 갈 수 있는 응급실을 알려줘.

The terminal should show location work as its own step, not hide it inside the final answer.

find

find retrieves public information. This is the live-heavy path: weather, emergency rooms, hospitals, road safety, welfare information, and other public-service reads.

Typical use:

다대포에서 김해공항까지 가야 하는데 비랑 도로 위험 정보를 같이 보고 판단해줘.

UMMAYA should only answer from results it actually found. If a live lookup returns one local hospital, the answer should not invent two more from another city.

check

check handles identity and scoped delegation. This is where Mobile ID, MyData consent, certificates, or OmniOne-style QR/app verification belongs: the user should scan or approve through the identity channel instead of typing raw personal data into the chat. In the current public package these identity tools are mock-backed or shape-mirrored unless a real provider credential exists. The hackathon path is to replace the mock identity ceremony with a live OmniOne CX / Open DID / Chain-backed check tool while preserving the same downstream delegation contract. A successful protected check returns a delegation context that downstream tools can consume.

Typical use:

모바일 신분증으로 본인확인하고, 내 자격으로 신청 가능한 복지 지원을 확인해줘.

The role of check is not generic advice. It is the protected-action boundary: identity proof, purpose, scope, consent, and delegation are established before UMMAYA moves toward a protected public-service action.

send

send executes or prepares submission, payment, application, receipt, or official handoff flows through wrapped public-service channels. It consumes the delegation context from check when the target domain requires identity or consent.

Typical use:

정부24 주민등록등본 발급 민원을 본인확인 후 신청 가능한 단계까지 진행해줘.

send is where UMMAYA moves beyond answer generation: if a live or permitted channel exists it calls that channel; if only a mock or official handoff is available it preserves the same receipt-shaped boundary and says so clearly.

Permissioned Work

UMMAYA separates public information from protected action.

Work typeUMMAYA behavior
Public lookupUse live public adapters when available.
Location resolutionUse live geocoding/address/region adapters when available.
Identity and delegationUse check to call the wrapped identity/consent channel, such as Mobile ID, MyData, certificates, or OmniOne-style QR/app verification.
Certificates, payment, submission, correctionUse send with the delegation context when a live or mock channel is registered; otherwise hand off to the official path.
API failure, zero result, missing authorityStop clearly instead of inventing a result.

This is not a separate safety pitch. It is part of how check and send work: the model plans the task, calls the right wrapped public-service tools, and stops at the authority boundary instead of acting like a reference-only RAG answerer.

Model And License

UMMAYA currently uses LGAI-EXAONE/K-EXAONE-236B-A23B through FriendliAI Serverless for LLM responses.

  • Model: K-EXAONE-236B-A23B
  • Reasoning mode: /reasoning or UMMAYA_K_EXAONE_REASONING_MODE selects fast, balanced, deep, diagnostic, or auto.
  • Thinking channel: UMMAYA_K_EXAONE_THINKING default false remains as a legacy compatibility flag; set /reasoning deep or UMMAYA_K_EXAONE_REASONING_MODE=deep for reasoning-channel diagnostics or benchmark runs.
  • Model license: K-EXAONE AI Model License Agreement
  • Project license: Apache License 2.0

The model license and the UMMAYA source-code license are separate. UMMAYA does not grant rights to LG AI Research, FriendliAI, Hugging Face, Korean government, or public-agency assets.

Ver o original no GitHub

Bom para

Relacionados