claudeers.
// Claude Plugins

HaiKrew

Token-efficient agent crews for Claude Code: Haiku writes, the main model specs and reviews, a live pixel crew pane shows who's working.

// Claude Plugins[ cli ][ api ][ desktop ][ web ][ claude ]#claude#plugins◷ MIT$open-sourceupdated about 1 month ago
Actively maintained
98/100
last commit 12 days ago
last release none
releases 0
open issues 0
// star history

Install with your AI

Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.

Install and set up HaiKrew (claude-plugin project) into my current project.
Found on https://claudeers.com/haikrew
Repo: https://github.com/adriancmurray/HaiKrew
Homepage/docs: —
Detected install method: claude-plugin → /plugin install haikrew@adriancmurray/HaiKrew
Category: plugins. Platforms: cli, api, desktop, web.
Read the repo's README for exact setup and env vars, then install it and wire it into my project.

Claudeers Health Verdict:
active; community-verified: false. Confirm the source before running anything.
// or install directly (claude-plugin)
/plugin marketplace add adriancmurray/HaiKrew
/plugin install haikrew@adriancmurray/HaiKrew
// or clone
git clone https://github.com/adriancmurray/HaiKrew

// compatibility

Platformscli, api, desktop, web
Operating systems—
AI compatibilityclaude
LicenseMIT
Pricingopen-source
LanguageTypeScript

Get your FREE $2.50 API credits to access TickAtlas financial data ↗

HaiKrew

HaiKrew: a crew of small orange robots under a pixel-art title

HaiKrew is a Claude Code mod (plugin) for token-efficient agent crews. A subagent launched without a model inherits the main model, so by default every helper runs on the most expensive one. HaiKrew sends code writing and research to Haiku, keeps the main model on planning and review, and shortens noisy command output with a hook instead of a model.

Unofficial community plugin; not affiliated with Anthropic.

Install

Requirements:

  • Claude Code 2.1.287 or later (for mods and plugin test).
  • Node 22.18 or later for the CLI. It runs TypeScript directly and has no build step or dependencies.

From GitHub, inside Claude Code:

/plugin marketplace add adriancmurray/HaiKrew
/plugin install haikrew@haikrew

The marketplace also lists clefkit, an optional decision backend (see Clefkit).

To try a clone for one session:

claude --plugin-dir /path/to/HaiKrew/plugin

Update

/plugin marketplace update haikrew

Then restart Claude Code. From a terminal, claude plugin update haikrew@haikrew does the same. The pane's Update button runs that command when you press it.

Quick start

  1. Run /haikrew. The pane opens with the model gate switch in its header (hotkey g). The gate is off after install.
  2. Turn the gate on. From then on, subagents launched without a model are routed (see Model gate).
  3. Give the main model a task that means writing more than a few lines. With the gate on, it is asked to hand the spec to haiku-coder, and the Your crew card lists each working agent.

What you see:

  • The pane has cards for Controls, Your crew, Token use, Track record, Insights and Settings; each folds to one line. The terminal shows a plain three-tab version: Home, Ledger, Settings.
  • The band (desktop app, above the prompt) shows the crew as small Claude bots. The main agent stands larger, in its model's colour (Haiku purple, Sonnet teal, Opus orange), wearing a crown, wizard hat or cap. Hover a bot for its name and latest line; its ≡ (transcript) button opens the agent's messages in a HaiKrew pane.
  • Hide the band with band.show set to off.

How it works

Model gate

  • Off by default (gate.enabled). Turn it on with the header switch, g, or Settings.
  • A subagent launched with no model gets one: the first matching rule in gate.rules, else gate.default_model (Haiku).
  • A model the caller names is kept, unless gate.rules_override is on. Opus is refused unless gate.allow_opus is on; the refusal returns to the caller with the reason and the setting.
  • Pinned agent types (haiku-coder by default) keep their own model. Forks inherit the parent's model, so rules do not apply to them.
  • A rule's @low to @max suffix, or gate.default_effort, sets how hard the subagent thinks. The main conversation is never changed.
  • With gate.delegate on (the default), a short note in the main conversation's context asks it to hand spec-able work to the crew.

The crew

  • haiku-coder implements a tightly written spec, runs haikrew verify and reports in 120 words or fewer.
  • haiku-scout answers one research question with a short cited brief. It has web search and fetch, read-only.
  • haiku-checker is an adversarial checker. It checks each spec item, runs the checks and tries to break the change. It is read-only on the repo.
  • sonnet-reviewer reviews the risky hunks of a standard or deep change, runs the tests and ends with a VERDICT line.
  • Skills: crew-overseer is the workflow (spec, dispatch, review by tier, verify). bench shows how to measure setups or models on the bench in bench/.
  • Each subagent gets a name, from its tag line or a list, shown in the crew card and transcript.

Review tiers

When a finished agent changes files, it gets a tier and the main conversation gets a one-line note. No tier is set when review.enabled is off or nothing changed.

  • Light: only review.light_paths changed (docs, tests), fewer than review.light_lines lines, and tests not failing. The checker checks it; with review.light_reviewer set to inline, the main model glances at the diff and calls record_verdict instead.
  • Standard: anything else. sonnet-reviewer at review.standard_reviewer (Sonnet, medium effort).
  • Deep: a sensitive path, review.deep_lines or more lines, a job in review.deep_jobs, failing tests, or a model under 60% over 4 or more judged runs at that job and language. The reviewer is review.deep_reviewer (Sonnet, high effort). With review.deep_inline on, the main model reviews it itself when it is at least as strong.
  • For standard and deep, the note tells the main model to launch the reviewer with HAIKREW job=review name=<name>. Its VERDICT: PASS|FAIL for <name> line sets the verdict of the run it reviewed.

Checker

  • The checker is the default light reviewer (review.light_reviewer is checker). It lists its break cases on one BROKE line; a case whose result names a failure counts as a defect.
  • Defects are counted on the checked run. The Insights card and haikrew analyze show the catch rate per job, language and size.
  • After review.checker_skip_after (15) checked light runs of one job, language and size with no defect, the checker is skipped for that kind and the main model glances instead. Every review.checker_recheck_everyth (5) skipped run is checked again. A defect resets the skip. Setting either to 0 turns that part off.

Ledger and insights

  • One row per finished subagent: model, job, language, time, tokens and verdict. Tokens are new work (input, cache writes and output); cache reads are kept apart.
  • A run is judged by the last VERDICT: or STATUS: done|blocked line of its final message. Runs with neither are unjudged and left out of pass rates.
  • The newest ledger.raw_limit (200) rows are kept in full. Older rows fold into monthly averages, and those older than ledger.yearly_after_months (12) into yearly ones.
  • Track record suggests a stronger model for a tagged job when Haiku or Sonnet passes under 60% over 4 or more judged runs. It suggests a cheaper model when Sonnet or Opus passes 90% or more over 8 or more, unless the model below has 4 or more judged runs of that job that pass poorly. Untagged runs get an inferred job, marked "(inferred)", which gets no rule suggestions.
  • Insights picks the cheapest model that passes 80% or more over 8 or more judged runs for a job and language (Opus only when allowed). That bar is separate from the Track record thresholds above.
  • A suggested rule is added to gate.rules only after its button is pressed twice.
  • The Insights card and haikrew analyze show checked pass rates, late failures, escalations, tokens per accepted result and the checker catch rate.
  • The ledger and review packets live in the mod's store. The mod reads no credentials and writes no other files.

Decider (shadow mode)

  • decider.mode is off by default. In shadow, the decider asks the endpoint at decider.url which model, effort and review a launch needs, and records the answer without acting on it. decide applies a confident answer; Opus still needs gate.allow_opus.
  • It runs only while the model gate is on (gate.enabled), and only for spawns the gate left undecided: no model named by the caller, no pinned type and no matching rule.
  • It sends the first 4000 characters of the task, the HAIKREW tag, the session folder's name and a table of past results to <url>/v1/systemone. Errors and timeouts leave the launch as the gate set it.
  • While the mode is on, <url>/health is checked. A failed check shows a warning on the Decider settings card.

Late failures

  • A run that passed is flagged late_fail when, within ledger.late_fail_days (5), a later commit in the same repo touches one of its files and its message says fix, revert, bug, broke or regress.
  • haikrew late-failures [--days N] writes the flags to ~/.claude/plugins/store/haikrew-late.json. A SessionStart hook runs it in the background.
  • The mod merges that file where it cannot run git itself, which is the desktop app. Without the file, the Insights card says detection is off.

Update notice

  • At most once every 6 hours, at session start, on the sky timer or on the first pane render, the mod reads version from plugin.json on GitHub. That is one request; nothing is sent.
  • When a newer version exists, the pane shows a strip with an Update button, and the main agent says so in a speech bubble in the band.
  • The button runs claude plugin update haikrew@haikrew when pressed, then asks for a restart. Turn the check off with pane.update_check.

Settings

Settings are stored in the mod's store and apply to every session. Edit them in /haikrew, where rules are built from dropdowns and validated on save. Defaults are in plugin/src/schema.ts.

NameDefaultWhat it does
gate.enabledfalseTurn the model gate on.
gate.default_modelhaikuModel for a subagent launched with none, when no rule matches.
gate.rulesjob:review=sonnet, job:research=haiku, type:Explore=haikuOrdered rules, first match wins: selectors=model[@effort]. Selectors: job:, lang:, type:, desc:, cwd:.
gate.allow_opusfalseAllow Opus subagents at all.
gate.rules_overridefalseLet rules replace a model the caller named.
gate.pinned_typeshaiku-coderAgent types that keep their own model.
gate.default_effortemptyEffort per model when no rule sets one, e.g. sonnet=high.
gate.allow_nestedtrueLet a subagent start agents of its own.
gate.delegatetrueAdd the delegation note to the main conversation's context.
gate.instructionsemptyExtra prompt text for matching subagents, as <selectors>: <text>.
decider.modeoffoff, shadow (record only) or decide (apply a confident pick).
decider.urlhttp://127.0.0.1:8790Base URL of the decision endpoint.
review.enabledtrueGive each finished agent that changed files a review tier.
review.light_reviewercheckerchecker (haiku-checker) or inline (the main model glances).
review.light_lines60Light reviews need fewer changed lines than this.
review.deep_lines400Changes of at least this many lines get a deep review.
review.standard_reviewersonnet@mediumReviewer model and effort for standard reviews.
review.deep_reviewersonnet@highReviewer model and effort for deep reviews.
review.deep_inlinefalseThe main model reviews deep changes itself when at least as strong.
review.checker_skip_after15Skip the checker for a light kind after this many runs with no defect. 0 never skips.
review.checker_recheck_every5While skipped, check every Nth light run. 0 never re-checks.
checks.guard_hookstrueRefuse Bash commands that skip git hooks (--no-verify, core.hooksPath).
checks.seam_checkfalseAlso submit the seam reminder as a prompt when parallel agents finish.
checks.token_limitshaiku=800000, sonnet=400000, opus=200000Warn when one subagent's new work passes its model's limit.
checks.token_effort_scaletrueScale token limits by effort, from 0.5x (low) to 3x (max).
squeeze.enabledtrueRoute matching build and test commands through haikrew squeeze.
squeeze.max_lines20Lines a squeezed summary prints.
read_guard.max_lines600Whole-file reads of longer files return an outline and ask for a range.
ledger.enabledtrueRecord finished subagents.
ledger.raw_limit200Newest rows kept in full.
ledger.late_fail_days5Window for late-failure flags.
agents.controloffWhat agents may do through HaiKrew's tools: off, read or change. Agents can never change this setting or Allow Opus.
pane.update_checktrueCheck GitHub every 6 hours for a newer version.
pane.open_on_startfalseOpen the pane when a session starts.
band.showtrueShow the crew above the prompt (desktop app).
band.skylivelive (your clock), dawn, day, dusk or night.
band.scenerymountainsmountains, city, forest or hills.
band.headwearcrownThe main agent's headwear: crown, wizard, cap or none.
band.haiku_color, band.sonnet_color, band.opus_color#9b7fc7, #4f9c98, #d97757Each model's colour in the band.

The table lists the main settings. The full set is in the settings field of the status agent tool's answer (agent tools on) and on the Settings card in /haikrew; gate.nested_model, decider.min_runs, squeeze.patterns and review.deep_jobs are among the rest.

Example: "job:implement+lang:rust=sonnet@high" runs Rust implementation specs on Sonnet at high effort. "cwd:~/projects/*=haiku" matches any session under ~/projects. A rule that picks Opus is refused unless gate.allow_opus is on.

CLI

The CLI runs anywhere Node 22.18 or later runs. The examples use haikrew, the launcher in plugin/bin/; node plugin/bin/haikrew.mjs is the same.

CommandWhat it does
haikrew squeeze [--max-lines N] [--wrap PREFIX] -- <cmd...>Run a command, print a short verdict and keep the full log under haikrew/logs in the system temp folder.
haikrew verifyRun .haikrew/verify if present, or an inferred build and test. Prints one VERDICT: line.
haikrew patterns check [--dir PATH] [--max-files N] [--max-lines N]Check a folder of .md pattern files against the caps (20 files, 40 lines). Exit 0 within caps, 1 on violations, 2 if the folder is missing.
haikrew analyze [--store PATH]Print the routing analysis shown in the Insights card, as a table.
haikrew late-failures [--days N]Flag late failures in the newest ledger store file (window 5 days by default) and write haikrew-late.json.

Clefkit (companion)

Clefkit provides typed decisions from Clef models and is the usual backend for the decider.

  1. Install it: /plugin install clefkit@haikrew.
  2. Start a backend as its README describes: Cloudflare Workers AI through a local wrangler proxy, a SystemOne or Jev URL, or a local server.
  3. Set decider.url to the backend's base URL, and decider.mode to shadow. The decider only asks while the model gate is on, and only for launches the gate left undecided.

The clefkit entry uses a git-subdir source with path plugin. That install path has not been tried against the published repository.

Troubleshooting

  • A subagent ran on the main model. It named a model of its own (kept unless gate.rules_override is on), it is a pinned type, or it is a fork, which inherits the parent's model.
  • Opus is refused. Turn on gate.allow_opus, or use a rule that picks Sonnet.
  • The desktop app shows an old version. It runs the copy cached at install. Update (see Update), then restart the app.
  • No band, or no pane. The band is desktop-only and needs band.show. The terminal shows the three-tab pane.
  • Insights says late-failure detection is off. Run haikrew late-failures from a terminal. It needs at least one ledger store file, which exists once a subagent has finished, and git in the repo.
  • The Settings card warns about the decider. The endpoint at decider.url did not answer <url>/health. Start the backend.

Known limits:

  • Reviewer effort is advisory. The review note names it, but HaiKrew cannot force a reviewer to that effort.
  • The gate has had one live check: an agent launched with no model ran on Haiku. Rules, effort, instructions and the seam check are covered by mod tests only.
  • The read guard and squeeze are heuristics. They can truncate output a caller needed; a squeezed run's full log stays on disk.

License

MIT. See LICENSE.

// faq

What is HaiKrew?

Token-efficient agent crews for Claude Code: Haiku writes, the main model specs and reviews, a live pixel crew pane shows who's working.. It is open-source on GitHub.

Is HaiKrew free to use?

HaiKrew is open-source under the MIT license, so it is free to use.

What category does HaiKrew belong to?

HaiKrew is listed under plugins in the Claudeers registry of Claude-compatible tools.

16 views
★ 11 stars
unclaimed
updated about 1 month ago

// embed badge

HaiKrew on Claudeers
[![Claudeers](https://claudeers.com/api/badge/haikrew.svg)](https://claudeers.com/haikrew)

// retro hit counter

HaiKrew hit counter
[![Hits](https://claudeers.com/api/counter/haikrew.svg)](https://claudeers.com/haikrew)

// reviews

// guestbook

0/500

// related in Claude Plugins

🔓

A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.

// pluginsmultica-ai/★ 230,631[ claude ]
🔓

Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explainin…

// pluginsanthropics/⟨Python⟩★ 156,472[ claude ]
🔓

"CLI-Anything: Making ALL Software Agent-Native" -- CLI-Hub: https://clianything.cc/

// pluginsHKUDS/⟨Python⟩★ 58,467◷ Apache-2.0[ claude ]
🔓

financial-services — a Claude ecosystem project on GitHub.

// pluginsanthropics/⟨Python⟩★ 41,409◷ Apache-2.0[ claude ]
→ see how HaiKrew connects across the ecosystem