Korean-first AI coding agent
Mirr Code
A Korean-first AI coding agent
A coding agent you talk to in the terminal, in Korean. It runs the tools and stays with the task until it is done.
Mirr Code runs on a model Patty built.
Mirr Code runs on models that Patty trains and serves itself. There is no external model API behind it, so Korean is not a translation target but the standard for training and evaluation.
Cursor movement during Hangul composition, initial-consonant search, reading repositories that mix Korean and English, and reading requirements out of public-sector documents all happen with no separate translation step.
See the Mirr modelJust say what you want done.
I want to automate the Excel work I do every day. Start with a plan.
- 01Understand the Korean request
Work out the screens and behavior you want from the request itself.
- 02Check the repository and team rules
Read the project structure, existing code, naming conventions, and team guidance to decide what to change and how.
- 03Plan the work and confirm authority
Propose which files change and why before touching them; anything that needs approval waits for it.
- 04Change and verify the code
Edit the code, then run type checks and tests in the same session. When something fails, Mirr finds the cause and fixes it.
- 05Summarize the result
Lay out the files changed, the checks that ran, and anything still worth a look.
Mirr is designed for the Korean working environment.
From input and search to document understanding and result explanations, the whole development workflow in Korean is taken into account.
Natural Hangul input
A syllable still being composed is never read as finished input. Cursor movement and deletion during Hangul input are tested continuously.
Search by initial consonants
Commands can be found by their Korean name, their English name, or their initial consonants alone. When several match, Mirr shows the candidates instead of picking one.
HWP and HWPX documents, understood right away
Mirr reads the body text, tables, and outlines of HWP and HWPX documents to find what the work needs. An encrypted document is reported as unreadable rather than guessed at.
Korean expressions in the repository, kept as-is
In a repository that mixes Korean and English, Mirr picks up the terms and notation already in use and follows them.
How Mirr works
Check the information needed, plan the change, then execute and verify within defined permissions.
- 01repository scan · HWP/HWPX reading · project guidance
Check what the work requires
Read the code, configuration, documents, repository guidance, and the current state of the work to identify what is needed.
- 02plan · affected files · validation path
Plan the change
Break the goal into steps and show which files will change and how they will be verified before execution.
- 03permission mode · sandbox · approval request
Execute within defined permissions
File edits and tool execution happen only inside the allowed permissions and sandbox. Work that needs additional authority requests approval first.
- 04checkpoint · rewind · error recovery
Revert when problems arise
When something goes wrong mid-task, Mirr returns to the previous state and runs again, checking test results to fix the cause.
- 05diff · test · DARI receipt
Leave results you can check
What changed, the checks that ran, the outcome, and what still needs review are kept as a record.
Shift+Tab decides how much you hand over.
The mode can change mid-session and applies from the next tool call.
Plan
Read-only. Nothing changes; Mirr only works out what it would do. Read the plan, then switch modes.
Normal
Every tool call is approved as it happens. Mirr asks before each file edit and each shell command.
Auto
Writes are delegated. The work runs inside the sandbox and stays there. Anything that has to cross that boundary still comes back for approval.
YOLO (Ctrl+Y)
Mirr runs to the end without asking. Use it for long repetitive work you do not want to sit through. Questions and plan approvals still reach you, and the checkpoints remain, so the session can be rewound.
| switch with Shift+Tab | read | write | shell |
|---|---|---|---|
| planRead-only. Produces a plan, changes nothing. | yes | blocked | trusted only |
| normalEach tool call is approved as it happens. | yes | approve | approve |
| autoWrites are delegated, inside the sandbox. | yes | yes | yes |
YOLO (Ctrl+Y) still stops for questions and plan approvals
Commands have Korean names.
There are 41 slash commands. The Korean name, the English name, and the initial consonants all find the same command.
| Korean | English | Choseong |
|---|---|---|
| /스킬설정 | /skills | ㅅㅋㅅㅈ |
| /샌드박스 | /sandbox | ㅅㄷㅂㅅ |
| /사용량 | /usage | ㅅㅇㄹ |
41 commands · an ambiguous entry lists candidates instead of running
- Rewind
- /돌아가기 · /rewind
※ The initials ㄷㅇㄱㄱ find it too.
- Usage
- /사용량 · /usage
- Skills
- /스킬설정 · /skills
- Korean compliance
- /컴플라이언스 · /pipa · /kisa · /csap
- Sandbox
- /샌드박스 · /sandbox
- All commands
- 41
※ Press / to open the palette. An ambiguous entry lists candidates instead of running one.
With other tools these meant a converter, or giving up.
The things Korean teams need, built in.
HWP and HWPX documents
Body text, tables, and outlines are read straight into the working context. No separate converter. An encrypted file is reported as unreadable rather than guessed at.
PIPA, KISA, CSAP checks
A single /컴플라이언스 command runs the review from the PIPA, KISA, and CSAP angles. /pipa, /kisa, and /csap all reach the same command.
Initial-consonant search
Commands and files are both searchable by initial consonants. Typing ㅁㄷ is enough to reach the model commands.
Hangul input itself
A syllable still being composed is not mistaken for a finished one. Cursor movement and deletion have their own path and regression tests.
How Mirr is built
See how Mirr is structured and the environments it runs in.
Configuration lives in markdown
Write project rules into MIRR.md, AGENTS.md, or CLAUDE.md and Mirr reads them at the start of every session. One @path line pulls in another file, and #note mid-conversation adds a rule on the spot.
Skills and plugins in one place
Installing one plugin brings skills, agents, slash commands, hooks, MCP tools, and themes as a set. You see what is inside before installing, and can disable, enable, or diagnose it afterwards.
Terminal first
The CLI, the TUI, and headless execution (mirr run) ship first. Desktop, ACP editor integration, and the HTTP server are in progress on the same runtime.
Our own model, and extensions
Patty develops and operates the model. Plugins, skills, MCP, agents, and ACP connections all run in the same runtime.
Why permissions and recovery are built in
Permissions are scoped so the agent works only where the task needs it to. Pre-execution confirmation, sandboxing, checkpoints, and rollback are part of the default flow, not an add-on.
Two personal plans, by how you work.
Prices are read live from the checkout catalog.
Pro
The standard plan for steady daily use
₩49,000₩29,000/ month
- Weekly and 5-hour usage allowances
- 2 parallel sessions
- HWP and HWPX document handling included
- Every runtime (CLI, desktop, headless)
Max
For long, unlimited sessions
₩99,000₩49,000/ month
- No usage limits
- 3 parallel sessions
- Long-running sessions resume from checkpoints
- HWP and HWPX document handling included
Start using Mirr Code.
Ask in Korean from your terminal, and the work begins.