Parts that keep Claude Code running by your rules.
Claude Harness Kit
An operations harness for Claude Code (24 parts)
The more you use Claude Code, the more cracks appear outside the tool itself. Rules skipped in long conversations. Sessions that quietly swell. Quota that drains without you noticing. Resident jobs that stop, and no one notices.
Claude Harness Kit is 24 parts that catch those cracks: a gate that blocks progress until the required rule has been read, a way to hand work to another window, measured conversation size and quota, and watchers for resident jobs. It is generalized from the setup Emocute runs every day in its own production environment, with anything tied to one person's circumstances removed.
Claude Code builds it itself. Have it read the setup guide and it asks which parts to install, then carries out placement through verification.
Who it's for
- Rules you wrote into CLAUDE.md stop being followed once sessions run long
- You want research, implementation and review handed to parallel Claude windows
- Multiple accounts or overnight automation make quota drain without you noticing
- You want to run Claude Code on your home Mac from outside
A day with it
- Build it — In the unpacked folder, say "Read CLAUDE_SETUP.md and build this kit." Claude asks which parts to install and finishes placement, configuration and verification.
- Rules hold — Ask it to reply to an email, and the gate sends the write back until the mail rule has been read. Adding a rule is one line in a table.
- See the size — A measured count of the conversation's size appears on the last line of every turn and at the bottom of the screen. Decide when to cut with a number, not a guess.
- Split windows — Independent work goes to another window with mado. A session that has grown heavy folds into a handoff with one word, tsugi, and continues in a new window.
- Guard the quota — claude-usage lists usage per account; claude-pace stops background automation once it is over pace.
- Away from the desk — Write in your own Discord channel and Claude Code runs on your home Mac, with progress relayed back to you.
- Watched — A watchtower checks resident jobs every 30 minutes and only notifies you when one has stopped. Silent when everything is fine.
How it works
- Start Claude Code in the unpacked folder and say: "Read CLAUDE_SETUP.md and build this kit."
- Claude asks which parts to install, then follows each part's guide through placement, configuration and verification.
- There is no settings screen. To change a behavior, tell Claude and it rewrites the files.
- Every part lives under ~/.claude/harness/ and ~/.local/bin/, and removal steps are included.
Complete kit
Claude Harness Kit 1.0 (24 parts)
$24
Buying every part separately comes to $150 (23 parts, excluding tailscale).
Parts
Every part is also sold on its own. Each line reads: the problem → what changes.
Keeping rules in force
Rules in CLAUDE.md get skipped in long sessions → Writes, sends and turn endings are sent back until the required rule has been read
Sessions quietly grow and get expensive → The measured token count appears on the last line of every turn
Uncommitted piles, bloated resident files, billing commands and degrading sessions go unnoticed → One-line notices only when something is off; only the dangerous ones are blocked (pick from five)
You restate how deep to dig every time → Fix a lean / balanced / rich default with one word
Commit style and rule writing drift between sessions → Two skill templates: commit-push and write-rules
Running many sessions
One session gets stuffed with everything and muddied → Hand work to Claude in another window and receive the report as markdown
You summarize heavy sessions by hand and paste them into a new window → One word writes a handoff and continues in a fresh window
You cannot tell what other windows are doing → A list of live sessions and a one-line roster in each conversation
Sessions pile up as "folder-0118" and cannot be found → A short title is taken from the first message and used as the window name
Windows stack on top of each other → New windows are placed into free slots automatically; windows you placed stay put
A Terminal crash takes every session window with it → Sessions run inside tmux and come back, windows and all
You cannot find where an old conversation was → Search every past session from a single page
Heavy work and small talk burn the same model's quota → Launch with a model chosen by how heavy the task is
Quota and the CLI itself
You lose track of which account is used where, and overnight automation eats quota → One entry point, usage readouts and a pacing gate
You do not notice a session growing → The measured token count stays on the bottom line
Missed updates and hard-coded model IDs keep you off new models → Updates every morning and rewrites pinned model IDs to aliases
Work during quota lockouts, or work that must stay local → Same interface, backed by a local Ollama model
Away from the desk
Claude Code only runs when you are at the Mac → Write in a channel on your own Discord server and it runs on your Mac at home, with live progress
You cannot reach Claude on your home Mac from outside → How to set up a private network with ssh / Remote Control (guide)
Ideas from outside or during lockouts get lost → They are delivered first thing in the next session
You miss permission prompts and finished work → macOS notifications tell you
Claude in Chrome stalls with "not connected" → The sleeping helper is woken automatically
Watching and tidying
Resident jobs stop without anyone noticing → A watch table is checked every 30 minutes, with notifications
Changes across several repos go uncommitted → Only tracked changes are committed, grouped by kind (push only when asked)
FAQ
Do I have to install every part?
Each part runs on its own. When you build the kit, Claude asks which parts to install, so you pick only what you need. Parts are also sold individually.
Is it worth it with just one account?
Yes. The gate, measured usage, window handoffs and watchtower all work regardless of how many accounts you have. accounts itself is useful with a single account, as a usage list and pacing gate.
What if something doesn't fit my setup?
There is no settings screen. Tell Claude what to change and it rewrites the files. Adjusting to your environment is up to you.
How do I remove it?
Each part's SETUP.md has removal steps. Everything lives under ~/.claude/harness/ and ~/.local/bin/.
Requirements
- macOS 14 or later / Claude Code (CLI) / python3 and zsh
- narabe needs Xcode Command Line Tools, fukki needs tmux, gemma needs Ollama, atelier needs a Discord account (a personal server is free)
- gate, ctxline, inbox and some guards also run on Linux
- Assumes a paid Claude plan (Pro or higher)
Notes
- Usage readouts (accounts) rely on an endpoint not described in Claude Code's public documentation. They may stop working if it changes; the other parts keep working.
- Claude Code updates may change how hooks behave.
- atelier runs through a Discord bot you create yourself. The bot token stays on your machine only.
- The setup guides and part documentation are in Japanese only.
- Adjusting to your environment is up to you. To change a behavior, tell Claude and it rewrites the files.
License
- Use in personal and commercial work environments
- Modify it for your own environment and use modified versions yourself or within your organization
- No redistribution, and no selling or publishing derivative products that include the kit
- The bundled LICENSE.md is authoritative
Digital download (zip)
Contact: [email protected]