Model or dataset
yohey-w/multi-agent-shogun avatar
yohey-w/multi-agent-shogun

The author of multi-agent-shogun gave up on his own army

Samurai-inspired multi-agent system for Claude Code. Orchestrate parallel AI tasks via tmux with shogun → karo → ashigaru hierarchy.

1,421 stars292 forksShellMIT

At a glance

What is it?
multi-agent-shogun runs coding agents in tmux panes under a Shogun, Karo and Ashigaru hierarchy, and coordinates them through YAML files on disk. Its own documentation now recommends a different project, and its test suite covers one of the seven command line tools it advertises.
Who is it for?
multi-agent-shogun earns its place if you want to watch named agents work in panes you can read, and if you already hold flat rate CLI subscriptions that make eight parallel sessions cheap. Two things to check before you commit.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 101 days ago.
What is it written in?
Mainly Shell, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • yohey-w/multi-agent-shogun1,421 ★ · +1
Star history
Repository2026-09-102026-11-15
yohey-w/multi-agent-shogun1.4k1.4k

1 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The sequel note sits above the quick start

The first block a reader meets is a note dated August 2026 saying the author has since disbanded his own ten-agent army down to a single agent that carries his judgment. The successor project, kagemusha, ships what the note calls the forms of a judgment loop, corrections turning into principles turning into standing rules, rather than the agents themselves, and links a Japanese write-up about the change. The note ends by saying multi-agent-shogun continues to work as is. That last sentence is the honest part of the document. The release line stopped earlier than the commits do: v5.1.0, named Karo as Traffic Controller, shipped on 23 May 2026, two days after v5.0.0 added first-class OpenCode support, and the last push on record is 6 August 2026.

Seven CLIs, seven sets of quirks and models

The pitch is vendor neutrality, and the comparison table claims seven command line tools where the alternatives manage one. The per-tool table is where the neutrality gets real and uneven. Claude Code brings the memory server and file tools and defaults to Claude Sonnet 4.6. Codex runs sandboxed with structured output and a headless exec mode, defaults to gpt-5.3-codex, and takes a per-model flag. Copilot defaults to the same Sonnet and adds four built-in agents plus a delegate command. OpenCode shares AGENTS.md instructions, defines agents through a flag, resets context with a slash command, and needs a restart for model changes. Cursor auto-loads CLAUDE.md and cursor rules and has a flag that auto-runs. Kimi defaults to k2. The Antigravity row stops mid-sentence, so its model is never stated.

Integration tests run against one of the seven

The Makefile separates unit tests from integration tests, and the integration target prints that it is Claude only. It first refuses to run at all when the tests/integration directory is missing, and it requires bats, printing an error telling you to run the install-deps target first. Then it filters: the bats call excludes the copilot and codex tags. Three of the seven advertised CLIs are therefore never exercised by that suite, and the remaining four outside Claude are untested too. That gap sits awkwardly against the release history, since the headline of v5.0.0 was first-class OpenCode support, and OpenCode is exactly the kind of client whose launch behaviour differs enough to need a test. Unit tests are described as root-level plus a tests/unit directory and run with timing output.

The Makefile declares fewer targets than it defines

The build file opens with a .PHONY line naming test, build, lint, check, help, install-deps and clean. Two targets that exist further down are missing from it: the integration target and a combined target that runs both suites. GNU make treats an undeclared file target as needing a rebuild, so with those two names absent from .PHONY a file named test-int or test-all in the working directory would shadow the recipes. The default target is help, which prints the seven commands it believes exist, and the recipes themselves are careful about their preconditions. The build target checks for the instruction builder and prints a warning naming scripts/build_instructions.sh when it is absent, then carries on. The lint target runs shellcheck over the lib and scripts directories only, so the shell files sitting at the top level, including both launch scripts, sit outside what the linter sees.

One line of Python under a shell runtime

The project is classified as shell, and the quick start is a clone, a setup script, a PATH reload and a departure script. The dependency file holds exactly one entry, PyYAML at version 6 or newer, which tells you where the Python surface sits: the coordination layer reads and writes YAML files on disk. That choice is the mechanism behind the headline claim of no coordination overhead. Instructions, task breakdowns, reports and decisions are files rather than messages inside one process, so the only model calls are the work itself. The documentation makes a point of those files being plain enough to read, diff and put under version control, and the chain of command is described as event-driven with no polling.

Five lines of quick start over a wide tree

The quick start is short enough to copy in one go:

bash
git clone https://github.com/yohey-w/multi-agent-shogun
cd multi-agent-shogun
bash first_setup.sh                        # one-time setup: config, dependencies, MCP
source ~/.bashrc                           # reload PATH
claude --dangerously-skip-permissions      # first run only: OAuth + accept Bypass Permissions → /exit
bash shutsujin_departure.sh                # launch all agents

The repository behind it is not short. The top level carries per-tool adapter directories for Claude, Cursor and OpenCode, a gitmodules file, a Makefile, AGENTS.md and CLAUDE.md, a CHANGELOG, CONTRIBUTING and SECURITY files, a Windows installer, and directories for agents, skills, instructions, templates, lib, scripts, config, context, memory, reports, docs, tests, images, android and saytask. The clone line carries no submodule flag, and the phone and voice lines at the top of the page have matching directories in the tree without appearing anywhere in the launch sequence above.

Coordination is free, permission bypass is not

The economics argument is that eight Opus-grade agents cost about $100 an hour through an API and about $200 a month on flat rate CLI subscriptions, and the page tells you to spend without hesitation because the bill does not move. Those are the project's own figures, not measured ones. The other cost sits in the first run line: permission prompts are bypassed for the session so the agents can work unattended, which is also what makes the YAML file trail worth reading. Two files in the tree point the other way, a gitleaks configuration for scanning secrets and a security policy. Nothing in the launch sequence ties them to the runtime, and the visibility claim rests on the panes themselves, since a file written by an agent is only as trustworthy as the agent that wrote it.

Editorial conclusion

multi-agent-shogun earns its place if you want to watch named agents work in panes you can read, and if you already hold flat rate CLI subscriptions that make eight parallel sessions cheap. Two things to check before you commit. First, the maintainer ran this setup down to a single agent and now points at a different project, so treat the hierarchy as a technique to study rather than a system with a future. Second, the automation is unattended and permission bypassed, so keep it on throwaway work, and read what the YAML files say before any of it touches a repository you care about.

Frequently asked questions

How many AI agents does multi-agent-shogun run at once?

The documented layout is one Shogun pane that delegates to a Karo, which coordinates 7 Ashigaru workers and 1 Gunshi strategist, each in its own tmux pane.

Which coding CLIs does multi-agent-shogun support?

Seven: Claude Code, OpenAI Codex, GitHub Copilot, Kimi Code, OpenCode, Cursor and Antigravity. Claude Code and Copilot default to Claude Sonnet 4.6, Codex to gpt-5.3-codex and Kimi to k2.

What does multi-agent-shogun need before it will start?

tmux, bash 4 or newer, and at least one supported coding CLI. The quick start is a clone, a one-time setup script, a PATH reload, and the departure script that launches the agents.

How do the agents in multi-agent-shogun talk to each other?

Through YAML files on disk, so the only API calls are for the work itself, and every instruction, report and decision is a file that can be read, diffed and version-controlled.

What does multi-agent-shogun claim running eight agents costs?

The page puts eight Opus-grade agents at roughly $100 an hour through an API against about $200 a month on flat rate CLI subscriptions. Those are the project's own figures rather than measured ones.

Is there a successor project to multi-agent-shogun?

Yes. The README points at kagemusha, which the author started after disbanding his own multi-agent setup down to one agent, and which ships a judgment loop of corrections, principles and standing rules.

Official sources

  1. Issues
  2. License: MIT
  3. README
  4. Releases
  5. yohey-w/multi-agent-shogun on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/yohey-w-multi-agent-shogun.svg)](https://hysenlabs.com/projects/yohey-w-multi-agent-shogun)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/yohey-w-multi-agent-shogun.svg)](https://hysenlabs.com/projects/yohey-w-multi-agent-shogun)

Was this useful?