OneWave AI Toolkit

A coding, marketing and sales team in one Claude plugin: 12 specialist agents, 16 skills, 4 team commands, 9 Claude Code mods and a secret guard that stops Claude from leaking your API keys.

Install it once and you can plan and review code, write and test it, build a launch kit, research a prospect and draft the outreach, all from the same session.

Install

In Claude Code:

/plugin marketplace add OneWave-AI/onewave-toolkit
/plugin install onewave-toolkit@onewave-toolkit

Then try /onewave-toolkit:team plan and build a pricing page, or just describe a task and Claude will pick the right agent.

The team: 12 agents

Discipline	Agent	What it does	Can change files?
Coding	architect	Maps the codebase, proposes an approach, breaks work into ordered steps and flags risks	No, read-only
Coding	code-reviewer	Reviews a diff or files for correctness, readability and maintainability, ranked by severity	No, read-only
Coding	test-engineer	Writes, fixes and extends tests in your existing framework, and runs them	Yes
Coding	debugger	Reproduces a failure, proves the root cause and applies the smallest fix	Yes
Marketing	content-strategist	Plans topics, formats, channels and a calendar tied to a business goal	Writes plan files
Marketing	seo-specialist	Audits pages for search intent, titles, meta descriptions, headings, links and technical issues	No, read-only
Marketing	copywriter	Landing pages, headlines, ads and emails that are specific and on-brand	Writes copy files
Marketing	social-media-manager	Turns any source into platform-native posts for LinkedIn, X and others	Writes post files
Sales	prospect-researcher	Builds a public-source brief on a company or contact, with every fact cited	Writes brief files
Sales	outreach-writer	Drafts personalized cold emails and follow-up sequences. It drafts, never sends	Writes draft files
Sales	deal-coach	Finds qualification gaps and risks in a deal and plans the next move	Writes notes
Sales	proposal-writer	Turns call notes into a proposal or recap with scope, timeline and next steps	Writes proposal files

Agents that research use web search. Agents that work on code use your normal Claude Code permissions, so you still approve anything your settings ask you to approve.

16 skills

Coding

• agent-army: run a 2-layer army of parallel agents on big refactors, migrations, audits and bulk generation
• scout: tells you which skill or agent fits the task you are on
• code-review-pro: a deep review covering security (OWASP Top 10), correctness, performance and maintainability
• test-coverage-improver: finds untested code and writes the missing tests
• plan-in-files: keeps a multi-step plan in files so long work survives a context reset
• session-handoff: writes a clean handoff when a session ends or work passes to someone else
• before-you-build: finds the biggest risk in a build before you start and the smallest way to test it

Strategy

• expert-debate-panel: a moderated debate between 4 or 5 expert roles with conflicting views, ending in a decision frame instead of a verdict. Delivered in chat, as a document or as an interactive HTML page

Marketing

• landing-page-copywriter: full landing page copy using proven frameworks such as PAS, AIDA and StoryBrand
• seo-optimizer: keyword analysis, readability scoring, meta descriptions and competitor comparison
• content-repurposer: turns any piece into 8+ formats, from social posts and an email series to slides, a video script and a one-pager
• linkedin-post-optimizer: hooks that survive truncation, readable line breaks and a hashtag strategy for LinkedIn

Sales

• cold-email-sequence-generator: 7 to 14 email sequences with A/B subject lines and follow-up timing
• prospect-research-compiler: prospect intelligence from news, company websites, social profiles and financial data
• sales-call-prep-assistant: pre-call briefs with company news, stakeholder backgrounds and discovery questions
• objection-pattern-detector: mines lost-deal notes for recurring objections and builds a response playbook from won deals

4 commands

Command	What it runs
/onewave-toolkit:team <task>	Picks the smallest set of agents for the task, runs independent ones in parallel and combines the results
/onewave-toolkit:review [target]	code-reviewer and test-engineer on your current changes, with a merge verdict
/onewave-toolkit:launch <product>	content-strategist, then copywriter, seo-specialist and social-media-manager in parallel, combined into one launch kit
/onewave-toolkit:prospect <company>	prospect-researcher, then outreach-writer, then deal-coach: a cited brief, drafted emails and call prep

9 mods (Claude Code)

Mods add live panes, a band above the prompt, status lines and guards to Claude Code itself. They run in the terminal and in the Desktop app's Code tab. Each one has an on/off switch in /config under this plugin. The two that read your inbox or call a model start off.

Mod	Command	What it does	Default
Burn Meter	/burn	Session cost as a growing fire bar above the prompt, plan limits with reset times, per-turn cost	On
Launch Codes	/launch-codes	Holds rm -rf, force pushes, DROP through a SQL client, production deploys and piped install scripts until you type a one-time code	On
Session Wrapped	/wrapped	An animated recap of the session and a shareable PNG card	On
Boss Fight	/boss	Failing tests spawn a pixel boss; every run that fixes tests lands a hit	On
Code Pet	/pet	A pixel pet that eats on tool calls, gets sick on failures and sleeps when you stop	On
Agent Narrator	/narrate	Every agent step in plain English, with a time-saved counter	On
Agent Race	/race	Race several sessions on the same task on a live scoreboard	On
Inner Monologue	/monologue	A pane of Claude's dry inner thoughts about the session	Off
Inbox Alerts	/alerts	Gmail, Slack and Calendar alerts as toasts and a pane	Off

What the mods reach

Mod	Network	Runs processes	Files	Calls a model
Burn Meter, Boss Fight, Code Pet, Launch Codes	No	No	No	No
Session Wrapped	No	python3 to read your local Claude Code transcripts, base64 to write the PNG	Writes one PNG to ~/Desktop and a cache in ~/.cache/session-wrapped	No
Agent Race	No	No	Reads and writes ~/.claude/agent-race/<race>/	No
Agent Narrator	No	No	No	Only with /narrate smart: the tool name and short fields, never file contents
Inner Monologue	No	No	No	Yes, a summary of recent tool calls and the first 120 characters of each prompt, while its pane is open
Inbox Alerts	Through your claude.ai Gmail, Slack and Calendar connectors	No	No	/alerts triage sends alert snippets to Claude

No mod makes its own network requests. Model calls go only to the Claude model your session already uses.

What the mods do under the hood

Everything below is also visible in the source under mods/ and hooks/register.tsx.

• Programs they run. Only Session Wrapped runs programs: python3 mods/session-wrapped/scripts/usage.py to total your week and month from local Claude Code transcripts, and base64 -D -o <card>.png (or, where base64 has no -o, python3 mods/session-wrapped/scripts/write_b64.py <card>.png) to save the recap card. No mod runs a shell.

• Files they write. Session Wrapped writes ~/Desktop/claude-wrapped-<time>.png and a cache at ~/.cache/session-wrapped/. Agent Race writes one small JSON file per session under ~/.claude/agent-race/<race>/. No mod writes build, settings or instruction files.

• Tools they call. Only Inbox Alerts calls tools itself, and only while its switch is on: the read-only claude.ai connector tools Gmail search_threads, Slack slack_search_public_and_private and Google Calendar list_events, every two minutes. What they return is shown in its pane and toasts and is not sent anywhere.

• Prompts they submit. Only /alerts triage submits a prompt: your unread alert snippets and upcoming meeting titles, fenced as untrusted text, asking Claude to rank them and draft replies without sending anything.

• Prompts they read. Burn Meter, Code Pet, Agent Narrator, Inner Monologue and Session Wrapped hook prompt.submit to note that a turn started (Burn Meter records the cost so far; the others count turns or note the request). None of them changes your prompt, and only Inner Monologue sends any of it to the model, as described above.

• Tool calls they watch. Several mods watch tool calls to react to them (the pet eats, the boss takes damage, the narrator explains the step). No mod changes a tool's input. Launch Codes can stop a Bash command until you confirm it, and if its own check ever fails it blocks the command rather than letting it through.

• Download-and-run patterns. Launch Codes and Agent Narrator contain the text of commands like curl … | sh so they can recognize and describe them. They never run them.

• Launch Codes is a speed bump, not a security boundary. It checks Bash commands only. Keep your normal permissions and sandboxing on.

• Inbox Alerts handles text other people wrote. /alerts triage wraps every email and Slack snippet in an untrusted block and tells Claude never to send, reply or delete anything. That lowers prompt-injection risk; it does not remove it. Set your email and Slack member ID in /config so your own messages do not alert you.

• The mods are also published one by one, with source and tests, at OneWave-AI/claude-code-mods. That repository also has Sportscaster, spoken play-by-play of your session, which is not bundled here.

Secret guard (hook)

The toolkit includes one settings hook that runs before Claude uses Bash, Write or Edit. It blocks three things:

1. A real API key or private key in a command or in a file Claude is about to write. It recognizes the key formats of Anthropic, OpenAI, Stripe, GitHub, GitLab, Slack, AWS and Google, plus PEM private keys.
2. Printing a secret variable with echo, printf or printenv, when the variable's name contains KEY, TOKEN, SECRET, PASSWORD or CREDENTIALS. That would put the value in your transcript. Piping the variable into a program and checking whether it is set ([ -n "$NAME" ]) are allowed.
3. Dumping the whole environment with a bare env or printenv.

When it blocks something, Claude sees the reason and takes a safer route. The hook is one readable shell script with no dependencies. It reads only the tool call it is checking, writes nothing, never reads your environment or files, and makes no network requests. Its test suite runs with sh tests/secret-guard.test.sh.

Privacy and data

• The skills, agents and commands are instructions only. They ship no code.
• The code is the secret guard hook and the mods, both described above. Neither makes its own network requests.
• Research agents use web search through your Claude session. Nothing else leaves your machine except your conversation with Claude itself.
• The toolkit collects no analytics. Mods keep small values (your burn total, your pet, boss kills) in Claude Code's own plugin storage on your machine.
• Sales agents work from public sources and never send messages for you: every email and post is a draft you review.

Where it works

Surface	Skills	Agents	Commands	Secret guard hook	Mods
Claude Code (terminal, IDE and the Desktop app's Code tab)	Yes	Yes	Yes	Yes	Yes. Panes and bands draw in the terminal and Desktop app
Cowork	Yes	Yes	Yes, as /onewave-toolkit:<command>	Yes	No
Claude chat (web, desktop, mobile)	Yes	No	Claude applies them like skills	No	No

Support

• Questions: onewave-ai.com/contact
• Bugs and ideas: GitHub issues
• Security issues: see SECURITY.md

Built by OneWave AI, an Anthropic Partner Network firm that trains teams on Claude and builds with it. MIT licensed.