Mahjong Engines
critter · Riichi · Source available

Akochan

The engine behind the first wave of free riichi game reviewers. Akochan is a C++ riichi AI with an mjai client, released for non-commercial research use.

GitHub ↗
Built by
critter (critter-mj) 1
Rules
Four-player Japanese riichi 14
Approach
C++ engine; reviews rank moves by expected pt 3
Strength
No ladder rank; loses to Mortal in duplicate play 4
Access
Source available, non-commercial terms 2

Who built it

Akochan is published by the GitHub user critter (critter-mj), with the repository created in August 2019 16. It describes itself simply as “Artificial Intelligence for Japanese mahjong” 1.

Rules

Akochan plays four-player riichi. It ships an mjai client and manager, so it can plug into mjai game servers and tools 1.

How it works

The repository documents how to build and run akochan, not its algorithm 1. What is documented is how it is used in reviews: the original akochan-reviewer ranked each candidate move by its expected value in Tenhou-style placement points (pt), or by expected placement, with a configurable pt table 3. Unlike Mortal, akochan’s reviews also show a deal-in rate for each move 4.

Equim, who built akochan-reviewer, notes that akochan sometimes plays oddly, a known issue Equim could not fix 5.

Strength

Akochan has no published Tenhou rank. In duplicate matches run by the Mortal project, with Mortal in three seats, akochan averaged −5.67 pt per game against Mortal 3.1 over 13,152 games, and −5.03 pt against Mortal 1.0 over 23,432 games 4.

How to access it

Online. mjai.ekyu.moe still offers akochan reviews, queued and taking 10 to 60 minutes 7.

Build it. On Linux, install Boost, then build the AI library and the main program 1:

sudo apt-get install libboost-all-dev
cd ai_src && make -f Makefile_Linux
cd .. && make -f Makefile_Linux
./system.exe test 100 102   # self-play with seeds 100–101

Windows builds use the plain Makefiles after setting the Boost library name and processor count, and a macOS Makefile is included 1. Self-play seats and tactics are set in setup_match.json 1.

Licence. The terms (in Japanese) allow free use for private research and study, and non-commercial publication of akochan’s output with credit. Other uses, including commercial publishing, need the author’s permission. Non-commercial redistribution is allowed with credit if the AI code is unmodified; redistributing modified AI code is generally not allowed 2.

Equim reports that cheating bots based on akochan have been sold for Mahjong Soul 5.

Sources

  1. Akochan repository and README (GitHub, critter-mj/akochan)
  2. Akochan licence terms (利用規約, Japanese)
  3. akochan-reviewer README (mjai-reviewer v0 branch)
  4. Mortal documentation, “Strength” (duplicate match results)
  5. Equim, “Should I release the trained mahjong AI model?” (gist, updated 2022-08-19)
  6. GitHub API record for critter-mj/akochan (creation date)
  7. Mahjong AI Utilities (mjai.ekyu.moe), service page

Facts checked against these sources on 24 October 2026.