AIgent Operator Kit

Five focused Claude Code specialists. One small install.

Echo maps, Newton researches, Hypatia challenges, Iris specifies, and Lyra builds. The optional context-loader hook adds a relevant file map and priority excerpt when your prompt mentions a configured project topic.

These agents also run inside aigent-OS, the full open-source operator system they were built for.

MIT licensed. The five agents are plain Markdown; the optional hook uses Bash and Python 3.8+.

Install

Recommended: Claude Code plugin

/plugin marketplace add wrg32786/operator-kit
/plugin install operator-kit@operator-kit

Plugin agents are namespaced, for example operator-kit:echo. Use the plugin or the legacy installer, not both. After installation, start a new session or run /reload-plugins, then invoke one naturally or by name:

use operator-kit:echo to find every place we call the Stripe API

Legacy user-scope install

curl -fsSL https://raw.githubusercontent.com/wrg32786/operator-kit/main/install.sh | bash

The legacy installer places unscoped agents under ~/.claude/agents/operator-kit/, preserves existing configuration, backs up customized agent files before replacing them, and wires the optional hook only when Python 3.8+ is available.

Manual agents-only install from a local clone:

mkdir -p ~/.claude/agents/operator-kit
cp agents/*.md ~/.claude/agents/operator-kit/

How routing works

You do not need to memorize the roster. Claude Code matches your request and current context against each agent's description, then delegates when a specialist is a good fit. Operator Kit descriptions explicitly say use proactively for their clear trigger cases.

Automatic delegation is best-effort, not a hard router:

• Ask naturally — Claude usually chooses the matching specialist.
• Name the agent — a strong hint, such as “use Echo to trace the callers.”
• Type @ and select the agent — guarantees that specialist runs for the task.

Each agent also carries the display color used by its character identity: cyan Echo, blue Newton, red Hypatia, purple Iris, and green Lyra.

Use the smallest team that works

Use one specialist by default. Compose only when the task crosses roles.

Need	Use
Locate code or trace callers	Echo
Implement a clear, bounded change	Lyra
Repair an unclear bug	Echo → Lyra
Evaluate a current tool or approach	Newton → Hypatia
Design and build a visual surface	Iris → Lyra
Review the current diff	Claude Code's built-in /code-review
Commit, push, open a PR, or merge	Main session, only when explicitly requested

Meet the agents

The roster gives each specialist a distinct visual identity while the actual behavior stays in the plain Markdown agent files.

Agent	Visual identity	Job	Example task
Echo	Cyan scout	Read-only codebase reconnaissance	“Echo, trace every caller of createInvoice and return paths and lines.”
Newton	Navy-and-gold researcher	Current, cited research synthesis	“Newton, compare Drizzle and Prisma for this schema using primary sources.”
Hypatia	Plum-and-crimson critic	Adversarial decision review	“Hypatia, find the strongest reason not to run this migration.”
Iris	Violet-and-white designer	Visual specification	“Iris, specify the dashboard empty state: hierarchy, type, spacing, and motion.”
Lyra	Green-and-steel builder	Bounded implementation	“Lyra, implement this accepted spec and run the smallest relevant check.”

The boundaries are enforced in frontmatter, not just prose. Echo, Hypatia, and Iris receive only read tools. Newton receives read and web-research tools. Lyra alone receives write and shell tools.

See the original terminal handoff animation

Review and publish

Focused analysis and implementation can route to agents. Repository side effects remain with the main Claude Code session:

1. Implement: Lyra edits and verifies the working tree.
2. Review: run the built-in /code-review when you want a deliberate diff review. It is manual by design so a longer review does not spend time and tokens unexpectedly.
3. Publish: explicitly ask the main session to commit, push, or open a pull request.
4. Merge: explicitly approve the merge, or use GitHub auto-merge after required checks pass.

No Operator Kit agent silently commits, pushes, opens a pull request, enables auto-merge, or merges.

Context loader

The plugin includes a silent UserPromptSubmit hook. It does nothing until you add a keywords file to your project.

mkdir -p .claude
curl -fsSL \
  https://raw.githubusercontent.com/wrg32786/operator-kit/main/examples/sample-project-keywords.json \
  -o .claude/operator-kit-keywords.json

Then replace the sample paths with paths from your project:

{
  "auth": {
    "keywords": ["authentication", "auth flow", "login", "session"],
    "priority_file": "docs/auth.md",
    "files": [
      "docs/auth.md",
      "src/lib/session.ts"
    ]
  }
}

When a submitted prompt contains one of those whole words or phrases, the hook adds:

• The first 40 lines of priority_file.
• A map of the configured project paths.
• An instruction to read the relevant paths before making claims or changes.

It matches only the submitted prompt, rejects paths outside the project root, writes nothing into the project, and keeps debug logging off by default.

Privacy: the priority excerpt becomes part of the active Claude session. Do not configure secrets, credentials, private keys, .env files, or any file you would not intentionally send to your configured Claude provider.

Full configuration and troubleshooting: context-loader/install.md.

Critical rules

rules/post-compact-critical.md.template is a starter for invariants that must survive long sessions and context compaction.

mkdir -p .claude/rules
curl -fsSL \
  https://raw.githubusercontent.com/wrg32786/operator-kit/main/rules/post-compact-critical.md.template \
  -o .claude/rules/critical.md

Use it for production/test boundaries, architectural invariants, known footguns, and required verification—not general documentation.

Platform support

The five agents are platform-independent Markdown. The optional context loader requires bash and Python 3.8+.

Environment	Agents	Context loader
Linux	Yes	CI-verified
macOS	Yes	Supported; Bash + Python required
WSL	Yes	Supported; Bash + Python required
Native Windows with Git Bash	Yes	Supported when bash and Python are on PATH; not yet CI-verified
Native Windows without Bash	Yes	No

Uninstall

Plugin install:

claude plugin uninstall operator-kit@operator-kit
claude plugin marketplace remove operator-kit

Legacy install:

rm -rf ~/.claude/agents/operator-kit ~/.claude/hooks/operator-kit
rm -f ~/.claude/operator-kit-keywords.json

Then remove the UserPromptSubmit entry that references ~/.claude/hooks/operator-kit/auto-context-load.sh from ~/.claude/settings.json.

Repository

operator-kit/
├── .claude-plugin/  # plugin manifest + marketplace catalog
├── agents/          # echo · hypatia · iris · lyra · newton
├── assets/agents/   # visual roster for the five specialists
├── context-loader/  # hook wrapper, standard-library loader, guide, starter config
├── examples/        # filled-out project keywords example
├── hooks/           # native plugin hook registration
├── rules/           # critical project-rules template
├── tests/           # dependency-free regression check
└── install.sh       # legacy user-scope installer

Validate

python3 tests/smoke.py
claude plugin validate .

CI also installs the actual Claude Code CLI into a clean configuration directory, adds this repository as a local marketplace, installs the plugin, verifies its agent and hook inventory, executes the installed context loader, and uninstalls it.

See CHANGELOG.md for release notes.

License

MIT. See LICENSE. The five agents are self-contained; the optional context loader uses only Python’s standard library.

---

Built by The AIgent. A weekly digest on running Claude Code at scale: theaigent.xyz

Want the full operator system these agents run inside? aigent-OS — free and open source (MIT).