Agent Toolkit - Setup Guide¶

Source: softaworks/agent-toolkit via skills.sh - 136.7K combined installs across 44 indexed listings; new to this catalog in the Oct 22, 2026 sweep

GitHub: softaworks/agent-toolkit - 2,587 stars, MIT license, last pushed 2026-03-05 (branch main)

Category: Developer Tooling / Agent Workflow

Quality Tier: 🟢 Production - MIT, 2,587 stars, CI workflows in the repo, and 43 of the 44 indexed listings backed by a SKILL.md in the tree; one sampled skill (qa-test-planner) carries MEDIUM warnings from Socket and Gen Agent Trust Hub for an eval pattern in its helper scripts

Agent Toolkit is a curated collection of skills for AI coding agents. Skills are packaged instructions and scripts that extend agent capabilities across development, documentation, planning, and professional workflows. In practice that means one repo you can point an agent at when you want Mermaid diagrams, cleaner prose, QA plans, typed React components, tidier commits, or a handoff document that survives a context reset.

The collection is broad rather than deep: 44 indexed listings drawing 136.7K combined installs, backed by 2,587 GitHub stars and an MIT license. Most listings are a single SKILL.md with optional helper scripts, so the cost of trying one is close to zero and the cost of installing all of them is prompt budget you may not want to spend.

Installation¶

Install the whole collection through the skills CLI:

npx skills add softaworks/agent-toolkit

Or register the repo as a plugin marketplace inside Claude Code and pull only the skills you want:

# Register as a plugin marketplace inside Claude Code
/plugin marketplace add softaworks/agent-toolkit
/plugin install humanizer@agent-toolkit
/plugin install codex@agent-toolkit

There is no separate build step documented, and each skill ships as instructions plus, in some cases, helper scripts that carry their own dependencies. After install the skills live as directories under the agent's skills path, each with a SKILL.md that the agent loads when a request matches its trigger description.

What It Provides¶

43 listings at or above the 500-install threshold, with install counts as indexed on October 22, 2026.

Skill	Installs	Use For
mermaid-diagrams	5,109	Guides creation of software diagrams in Mermaid syntax: class, sequence, flowchart, and entity relationship visuals.
humanizer	4,408	Detects and fixes AI writing tells, including inflated symbolism, promotional language, and em dash overuse.
qa-test-planner	4,327	Generates test plans, manual test cases, regression suites, and bug reports, with Figma MCP integration for design validation.
database-schema-designer	4,325	Designs SQL and NoSQL schemas with normalization, indexing, migration, and constraint guidance.
writing-clearly-and-concisely	4,305	Applies Strunk's rules to documentation, commit messages, error messages, reports, and UI text.
skill-judge	4,231	Scores SKILL.md files and skill packages against official specifications, with improvement suggestions.
c4-architecture	4,219	Produces C4 architecture documentation as context, container, component, and deployment diagrams.
agent-md-refactor	4,213	Splits bloated AGENTS.md or CLAUDE.md files into linked docs using progressive disclosure.
session-handoff	4,188	Writes handoff documents for session transfers when context fills or work pauses.
crafting-effective-readmes	4,182	Templates and guidance for README files matched to audience and project type.
commit-work	4,147	Reviews and stages changes, splits logical commits, and writes Conventional Commits messages.
marp-slide	4,106	Builds Marp decks across seven themes, from default and minimal to dark, gradient, and business.
react-dev	4,066	Type-safe React 18 and 19 patterns: generic components, hook typing, event typing, and routing integration.
professional-communication	4,056	Email structure, team messaging etiquette, meeting agendas, and audience-adapted technical messages.
codex	4,055	Runs Codex CLI (codex exec, codex resume) for analysis, refactoring, and automated edits, GPT-5.2 by default.
dependency-updater	4,032	Auto-detects project type, applies safe updates, prompts on major versions, and diagnoses dependency issues.
command-creator	4,023	Builds Claude Code slash commands with proper structure and best-practice conventions.
react-useeffect	4,017	Official-docs useEffect guidance, including when not to use Effect and what to use instead.
meme-factory	4,011	Generates memes through the memegen.link API across 100+ templates with custom text and styling.
reducing-entropy	3,997	Manual-only skill that shrinks codebase size, biased toward deletion and measured by final code amount.
ship-learn-next	3,983	Turns transcripts, articles, and tutorials into implementation plans via the Ship-Learn-Next framework.
openapi-to-typescript	3,976	Converts OpenAPI 3.0 JSON or YAML into TypeScript interfaces and type guards.
gemini	3,967	Runs Gemini CLI for code review, plan review, and large-context analysis, Gemini 3 Pro by default.
feedback-mastery	3,957	Preparation-Delivery-Follow-up and Situation-Behavior-Impact frameworks for difficult conversations.
domain-name-brainstormer	3,943	Generates domain name ideas and checks availability across multiple TLDs, including .com, .io, .dev, and .ai.
daily-meeting-update	3,940	Four-question standup interview that pulls GitHub, Jira, and session history into a Markdown update.
difficult-workplace-conversations	3,938	Preparation-delivery-followup approach to conflicts, performance discussions, and critical feedback.
plugin-forge	3,937	Creates Claude Code plugins with manifests, components, version bumps, and marketplace wiring.
game-changing-features	3,924	Finds 10x product opportunities and high-leverage improvements through strategic product thinking.
naming-analyzer	3,898	Suggests better variable, function, and class names from context and conventions.
lesson-learned	2,983	Mines git history for engineering lessons and principles from recent code changes.
gepetto	1,540	Builds sectionized implementation plans through research, stakeholder interviews, and multi-LLM review.
requirements-clarity	946	Clarifies ambiguous requirements with a YAGNI check and a KISS check before coding.
jira	820	Issue lookup, creation, updates, and sprint or backlog workflow for Jira tickets.
draw-io	807	draw.io diagram creation and review: .drawio XML editing, PNG conversion, layout, and AWS icons.
backend-to-frontend-handoff-docs	791	Writes API handoff documentation for frontend integration work.
frontend-to-backend-requirements	788	Documents frontend data needs as API requirements for backend developers.
mui	787	Material-UI v7 patterns: sx prop styling, theme integration, responsive design, and MUI hooks.
design-system-starter	786	Design tokens, component architecture, accessibility guidelines, and documentation templates.
excalidraw	782	Delegates Excalidraw file work to subagents so verbose diagram JSON does not exhaust context.
perplexity	763	Web search and research for generic queries through Perplexity, not library docs or workspace questions.
web-to-markdown	751	Converts JS-rendered pages to Markdown through the local web2md CLI built on Puppeteer and Readability.
datadog-cli	709	Log search, metric queries, request tracing, and dashboard management from the Datadog CLI.

The remaining 1 indexed listing is lessons-learned (1 install), indexed on skills.sh but with no SKILL.md in the current repository tree.

Why This Matters for Hermes Agents¶

Hermes loads skills from a directory of SKILL.md files, which is the same shape this repo ships. You can drop selected directories into the profile's skills path and get the same trigger-based loading Hermes already uses, without a wrapper or a translation layer.

The collection fills gaps that come up constantly in agent work. commit-work and agent-md-refactor deal with the two artifacts that rot fastest in a long-lived repo: git history and the instruction file. session-handoff is the direct counterpart to context compression, and skill-judge gives you a scoring rubric for the skills you write yourself.

It is also a useful reference set. reducing-entropy is deliberately manual-only, so it will not fire unless asked, which makes it safe to keep installed. writing-clearly-and-concisely and humanizer cover different problems, one structural and one stylistic, and they compose well when you are publishing prose.

Two caveats worth internalizing. First, install counts measure attention, not maintenance. Second, the repo's own quality signal is uneven across 43 listings, and the sampled verdicts in the Security section show at least one that needs a read before you trust it with a shell.

Usage¶

After installation, invoke a skill by describing the task the way its trigger expects. The agent matches on the skill description, so the practical skill is naming the artifact you want:

"Diagram this service's request flow" routes to mermaid-diagrams or c4-architecture depending on whether you want a flowchart or a C4 view.
"Clean up the AI phrasing in this announcement" routes to humanizer.
"Write a test plan and regression suite for the checkout flow" routes to qa-test-planner.
"Split these staged changes into logical commits" routes to commit-work.
"Save a handoff before I close this session" routes to session-handoff.

Chaining works the way it does for any skill: produce the artifact with one skill, then hand it to another. A common sequence on documentation work is mermaid-diagrams for the diagram, writing-clearly-and-concisely for the surrounding prose, then humanizer for a final pass on anything customer-facing.

If you only want part of the collection, use the plugin marketplace path and install named skills rather than the whole repo. Keeping the installed set small matters: every installed skill adds a line to the agent's skill index, and a 43-skill install competes for attention with your own internal skills.

Verification¶

Confirm the repo layout before you rely on it. Each skill should resolve to a SKILL.md under dist/plugins/<skill>/skills/<skill>/SKILL.md. Three paths verified on October 22, 2026, all returning HTTP 200:

for s in mermaid-diagrams humanizer qa-test-planner; do
  curl -sL -o /dev/null -w "%{http_code} $s\n" \
    "https://raw.githubusercontent.com/softaworks/agent-toolkit/main/dist/plugins/$s/skills/$s/SKILL.md"
done

https://raw.githubusercontent.com/softaworks/agent-toolkit/main/dist/plugins/mermaid-diagrams/skills/mermaid-diagrams/SKILL.md
https://raw.githubusercontent.com/softaworks/agent-toolkit/main/dist/plugins/humanizer/skills/humanizer/SKILL.md
https://raw.githubusercontent.com/softaworks/agent-toolkit/main/dist/plugins/qa-test-planner/skills/qa-test-planner/SKILL.md

A 200 on all three means the tree still matches the documented layout and the install command has something to resolve. A 404 on a specific skill means that listing was renamed or removed since this snapshot, so trust the tree over the install counts rather than the other way around.

Security¶

Sampled verdicts from skills.sh, verified October 22, 2026:

Skill	Gen Agent Trust Hub	Socket	Snyk
mermaid-diagrams	Pass	Pass	Pass
humanizer	Pass	Pass	Pass
qa-test-planner	Warn	Warn	Pass
database-schema-designer	Pass	Pass	Pass
writing-clearly-and-concisely	Pass	Pass	Pass

The one warning is specific: qa-test-planner carries MEDIUM warnings from Socket and Gen Agent Trust Hub for an eval pattern in its helper scripts. That is not a verdict on the prose instructions, it is a flag on executable code, and it is the kind of thing that matters because a skill's scripts run with whatever permissions your agent already has.

Practical rules. Read the helper scripts of anything that shells out before you install it, especially eval patterns, network calls, and file writers. Install from the plugin marketplace a few skills at a time rather than pulling all 43 blind. Use skill-judge on any skill you plan to keep, including third-party ones, since it scores against the same specification the repo cites. And treat a Pass from a scanner as a floor, not a guarantee: five sampled verdicts out of 43 listings is a sample, not a full audit.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified October 22, 2026: 136734 combined installs across 44 indexed listings; 2,587 GitHub stars; license MIT; last pushed 2026-03-05. Counts drift over time.

44 indexed listings, but only 43 have a SKILL.md in the current tree. lessons-learned is indexed without a tree counterpart and sits below the install threshold at 1 install.
The repo has not been pushed since 2026-03-05, more than seven months before this snapshot. Skills that wrap fast-moving CLIs, notably codex, gemini, and datadog-cli, can drift from the current tool surface.
One sampled skill, qa-test-planner, carries MEDIUM warnings from Socket and Gen Agent Trust Hub for an eval pattern in its helper scripts. The other four sampled skills passed both scanners and Snyk.
Two pairs overlap heavily: feedback-mastery and difficult-workplace-conversations both cover a preparation-delivery-followup approach to hard conversations, and writing-clearly-and-concisely and humanizer both touch prose quality. Install counts alone will not tell you which one to pick.
reducing-entropy and web-to-markdown are manual-only by their own instructions. They will not trigger from ambient context, which is a feature but also a surprise if you expect automatic behavior.
Install counts are a point-in-time index, not a quality signal, and they cluster between 709 and 5,109, so rank order among mid-pack skills says very little.

Pro Tips¶

Install humanizer and writing-clearly-and-concisely together, then run the structural edit before the stylistic one. Doing it in the other order means you humanize prose you are about to rewrite.
Use agent-md-refactor when your AGENTS.md passes a few hundred lines. Progressive disclosure beats a longer instruction file every time, and the skill is opinionated about the split.
Run skill-judge on your own skills before you publish them. It scores against the same specification this repo follows, so its critique transfers.
Point session-handoff at a specific next action rather than a general summary. Handoffs that end in "continue the work" cost more context than they save.
For architecture work, start with c4-architecture for the container and component views, then use mermaid-diagrams only for the flows that need a sequence or flowchart. Two diagram skills on one document produce inconsistent notation.
Check the tree, not the install count, when a skill misbehaves. The three verified raw paths above are the fastest way to confirm a listing still exists.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ