softaworks/agent-toolkit: a curated skill pack for Claude Code

What softaworks/agent-toolkit actually is

This is not a framework and not a runtime. The README describes it as "opinionated skills shared by @leonardocouy for improving daily work efficiency with Claude Code," where a skill is a packaged set of instructions and scripts that extends what the agent can do. The repository is a distribution container for those packages, plus the agents and slash commands that ship alongside them.

The audience is narrow and specific: people who already work inside Claude Code and want prewritten behaviour for recurring tasks. The skill list reads like a slice of one developer's week. There are planning skills such as gepetto and requirements-clarity, documentation skills such as c4-architecture and crafting-effective-readmes, frontend skills such as react-dev and openapi-to-typescript, and a professional category covering standup updates, feedback delivery and difficult workplace conversations. A meta category exists too: skill-judge evaluates skill design, plugin-forge builds plugin manifests, and command-creator generates slash commands.

That breadth is the point and also the first thing to scrutinise. Nothing in the repository indicates a shared abstraction across these skills beyond the Agent Skills format at agentskills.io. Each one is its own plugin, installed separately, which means the toolkit is closer to a catalogue than a product.

How skills, agents and commands are packaged

The repository layout tells you most of the architecture: agents/, commands/, skills/ and dist/ sit at the top level, with .claude-plugin/ holding the marketplace metadata and scripts/ holding helper code. The primary language is Python, which matters because several skills wrap scripts rather than being pure prose instructions.

The README states that every skill, agent and command is an individual plugin that can be installed on its own. The naming convention encodes the type: skills install as their bare name, agents as agent-<name>@agent-toolkit, and commands as command-<name>@agent-toolkit. So codex and humanizer are skill plugins, agent-codebase-pattern-finder is an agent plugin, and command-codex-plan is a command plugin.

This is a flat model with no dependency resolution between entries. Installing humanizer does not pull in skill-judge, and there is no documented notion of a versioned bundle. Whether that is a flaw depends on what you want. For a single developer curating their own workflow, flat plugins mean you install only what you use. For a team trying to standardise on a shared set, it means every member assembles their own subset, and the README offers no manifest that pins the whole set at once.

Installing agent-toolkit and running a first skill

The README gives a quick install path that it recommends, using the skills CLI. It states this method works with multiple AI coding agents including Claude Code, Codex, Cursor and AdaL.

npx skills add softaworks/agent-toolkit

If you are working inside Claude Code and prefer the native plugin system, register the repository as a marketplace first, then open the plugin interface.

/plugin marketplace add softaworks/agent-toolkit
/plugin

Inside that interface, the README describes switching to the Marketplaces tab, selecting agent-toolkit, and browsing the plugins. You can also install a specific entry directly by name. The README's own examples install the codex skill and the humanizer skill.

/plugin install codex@agent-toolkit
/plugin install humanizer@agent-toolkit

There is a manual fallback for Claude Code that copies a single skill directory into your user skills folder. Replace the placeholder with a real directory name from the skills/ tree, for example codex.

cp -r skills/codex ~/.claude/skills/

Updating is a marketplace operation rather than a package upgrade: run /plugin, go to the Marketplaces tab, select agent-toolkit and choose Update marketplace. The README also notes that auto-update can be enabled. For claude.ai, the documented path is different again: add skills to project knowledge or paste the contents of SKILL.md into the conversation. That last option is worth pausing on. Pasting a SKILL.md by hand means the skill is not versioned anywhere, and the README does not describe how to keep such a copy in sync.

Where the toolkit stops being the right choice

The first limitation is the install surface. Everything documented assumes Claude Code, the skills CLI, or manual copying into ~/.claude/skills/. There is no package on a language registry, no server, no Docker image, and no configuration file for CI. If your team needs skills provisioned reproducibly across machines, the README does not describe a mechanism for that beyond the marketplace and auto-update toggle.

The second is the update story. The README documents updating plugins through the marketplace UI and enabling auto-update, but it does not document rollback, pinning a skill to a specific revision, or what happens to an installed skill when the upstream directory is renamed. Skills are referenced by path and name, so a rename in the repository is a breaking change for anyone who installed by that name.

The third is scope. Several skills are about writing and workplace communication rather than code: daily-meeting-update, feedback-mastery, difficult-workplace-conversations. That is deliberate, and the README is explicit that the collection covers professional workflows. But if you are looking for a coding agent toolkit in the infrastructure sense, agent orchestration, tool routing, model access, this repository does not provide it. It provides prompt-and-script packages that run inside an agent you already have.

Compared with writing your own SKILL.md files

The real alternative is not another repository. It is the Agent Skills format itself. Because skills follow the agentskills.io specification, you can write a SKILL.md for your own conventions and place it in ~/.claude/skills/ without depending on anyone else's curation.

The difference in approach is ownership of the content. A self-written skill encodes your project's actual conventions: your test command, your directory layout, your review checklist. A skill from this toolkit encodes one developer's opinion, which the README openly calls "opinionated." That is useful when your conventions are unformed and you want a starting point, and it is friction when they are already settled and the skill's instructions disagree with them.

There is a hybrid worth naming. The repository includes skill-judge for evaluating skill design and plugin-forge for building plugin manifests, so the toolkit can be used as a reference implementation while you write your own. You read how a skill is structured, then produce a version that matches your codebase. That path costs more time than installing codex@agent-toolkit, but it does not leave you maintaining a fork of someone else's prompt.

Maintenance, licensing and what to check before adopting

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the licence and copyright notice are retained. That is the permissive end of the spectrum and it removes the usual legal friction for internal adoption. It does not, however, settle attribution for the content of individual skills, and the LICENSE file at the top level is the document to read rather than this summary.

On maintenance, the honest position is thin. The repository is not archived, but the last push was on 2026-03-05, and no releases are published. There is no changelog in the repository and no version tags to pin against. Upgrades therefore happen through the marketplace update path, which moves you to whatever is on the default branch at that moment. The practical cost is that a skill's behaviour can change without a version number to notice.

Before adopting, check three things in the repository itself. Confirm the skill you want exists under skills/<skill-name>, since the README's table is the only index. Read that skill's own README, because the top-level document describes categories rather than behaviour. And decide whether the plugin marketplace or the manual copy into ~/.claude/skills/ is your install path, because they update differently and only one of them is covered by the marketplace update flow.

Editorial conclusion

Adopt it if you already run Claude Code and want a menu of ready-made skills for planning, documentation and frontend work without writing SKILL.md files from scratch. Skip it if you need a general agent framework, a hosted service, or a project with a documented release history: no releases are published, and the last push was on 2026-03-05. Before installing, open the skills directory and confirm the specific skill you want exists at skills/<skill-name>, because the README's table is the only index and the plugins are installed one at a time.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. softaworks/agent-toolkit on GitHub