UMMAYA

npm version

Terminal AI agent for Korean public-service workflows.

UMMAYA is an open-source civic AI agent that brings a Claude Code-style tool loop to Korean public-service and national administrative workflows. You ask for an outcome in natural language; UMMAYA routes the request, calls registered civic adapters, shows tool progress, and stops at identity, consent, payment, or official-authority boundaries.

If Codex and Claude Code are terminal agents for software work, UMMAYA is the same class of interface for Korean public-service work: a terminal AI agent, public-service tool-calling harness, government-workflow CLI, and national AX reference client powered by K-EXAONE on FriendliAI.

UMMAYA is an independent academic and R&D project. It is not affiliated with, endorsed by, or operated by Anthropic, LG AI Research, FriendliAI, the Korean government, or any public agency.

UMMAYA terminal demo showing public-service prompts, tool calls, tool results, and final answers

_{Click the demo to open the MP4. The recording is from a real ummaya terminal session; waiting time was trimmed, but prompts, UI, tool calls, and answers were not synthesized.}

Find This Repo When You Need

Search intent	Why UMMAYA matches
Terminal AI agent for Korean public services	Runs as a local CLI and answers Korean public-service requests through tool calls.
Civic AI agent or govtech agent	Wraps public-information, location, identity, submission, and document workflows behind auditable adapters.
Claude Code-style tool loop outside coding	Migrates the agent loop, permission UX, context assembly, and terminal interaction model from developer work to public-service work.
K-EXAONE or FriendliAI agent example	Uses `LGAI-EXAONE/K-EXAONE-236B-A23B` through FriendliAI Serverless as the fixed model path.
LLM tool calling for government workflows	Exposes a small action vocabulary while adapters carry agency-specific details and evidence.

Canonical discovery terms: terminal AI agent, Korean public-service workflows, civic AI agent, government workflow CLI, public-service tool calling, national AX, K-EXAONE, FriendliAI, Claude Code-style harness, llms.txt.

Install

npm install -g ummaya

Homebrew on macOS:

brew install --cask umyunsang/ummaya/ummaya

The Homebrew cask installs a prebuilt macOS archive. It 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.

First Run

ummaya

Inside the terminal session:

/login

Paste your FriendliAI API key, then ask in everyday language:

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

Public packaged CLI users should not need to prepare Kakao, JUSO, SGIS, data.go.kr, KMA/APIHub, gateway bearer, or other operator-managed public API credentials. Sessions are stored locally; resumable sessions print a command such as ummaya --resume <session-id>.

What You Can Ask

Ask for the public-service outcome, not the agency API name.

User goal	Example prompt	Expected boundary
Find nearby public information	`다대1동 근처에서 오늘 전화해볼 수 있는 내과가 있을까? 가까운 곳 위주로 주소랑 전화번호를 알려줘.`	Live public lookup when an adapter is available.
Combine weather, safety, and travel context	`비 오는 날 다대포에서 김해공항까지 차로 가야 해. 날씨랑 도로 위험 정보를 같이 보고 조심할 점을 알려줘.`	Live public data plus clear source boundaries.
Prepare a protected public-service workflow	`모바일 신분증으로 본인확인하고, 내 자격으로 신청 가능한 복지 지원을 확인해줘.`	`check` starts identity or delegation flow; Mock or Handoff if no live authority exists.
Draft a public form	`전입신고서 양식을 열어서 내 정보로 채워줘. 제출 전에 어떤 칸이 비었는지도 확인해줘.`	`document` inspects and prepares; saving requires evidence and approval.

How It Works

UMMAYA keeps the model-facing surface small. The LLM sees a few stable actions; adapters handle agency-specific APIs, schemas, credentials, mock fixtures, and evidence.

Layer	Role
Harness	Claude Code-style loop: plan, call tools, inspect results, continue or stop.
Model	`K-EXAONE` through FriendliAI Serverless.
Tool registry	Korean public-service adapters registered as model-callable tools.
Permission UX	Explicit consent and authority gates before protected actions.
Evidence	Status labels, receipts, source traces, and local session history.

Main Tool Families

Tool	Public alias	Role
`locate`	`resolve_location`	Resolve places, addresses, coordinates, and administrative areas before a service lookup.
`find`	`lookup`	Retrieve public information such as weather, emergency rooms, hospitals, road safety, welfare information, and other public-service reads.
`check`	`verify`	Enter protected work by establishing identity, consent, purpose, scope, and delegation. This is the light permission gate.
`send`	`submit`	Execute or prepare submissions, payments, applications, receipts, or official handoffs when a live or mock channel is registered. This is a heavy permission gate.
`document`	`document`	Inspect, fill, style, validate, render, and save Korean public forms such as HWP, HWPX, PDF, OOXML, and ODF. This is gated by Evidence Fabric coverage and explicit in-session approval.

Capability Status

UMMAYA is useful only if it is honest about authority. Every domain is treated as Live, Mock, or Handoff.

Status	Meaning
Live	An official callable channel, valid credential path, and documented contract are available.
Mock	The channel exists or is policy-mandated, but this repo lacks live authority; UMMAYA mirrors the public shape without pretending completion.
Handoff	Only an official portal, person, app, or protected rail can complete the work today.

Live public-information flows are the easiest to verify today, so demos focus on those. Identity, submission, payment, certificate, correction, and some form-writing paths are Live only when an official callable channel, valid credential, and documented contract exist.

For Users, Researchers, And Agents

Reader	Start here
End users	Quickstart and what you can ask
LLM research tools	llms.txt and LLM-readable docs
Adapter contributors	Adapter authoring and docs/api/
Architecture reviewers	docs/vision.md and docs/requirements/ummaya-migration-tree.md

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; use /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 UMMAYA source-code license are separate. UMMAYA does not grant rights to LG AI Research, FriendliAI, Hugging Face, Korean government, or public-agency assets.

Docs

English docs: ummaya-docs.pages.dev/en
Korean docs: ummaya-docs.pages.dev/ko
Agent-readable discovery: llms.txt
Architecture source of truth: docs/vision.md
Requirements tree: docs/requirements/ummaya-migration-tree.md
Adapter docs: docs/api/
Codex continuation: docs/onboarding/codex-continuation.md
Release readiness: docs/release/homebrew-official-readiness.md
Contributing: CONTRIBUTING.md