Claude Skills Marketplace: Plugins for Claude Code Workflows, and the Execution Runtime Behind Them

The gap this fills in a Claude Code setup

Claude Code ships with slash commands, which a person has to invoke by name. This repository is built around the opposite trigger: skills are model-invoked, and the README states they are "automatically triggered by Claude based on context and the Skill's description." That distinction is the whole product. If you want Claude to reach for a git workflow, a test-fixing routine or a code-review pass without you remembering the command name, you have to write the skill description carefully and let the model decide. The marketplace supplies those descriptions already written, for four domains: engineering workflow, visual documentation, productivity, and code operations. The audience is narrow and specific. It is engineers who already use Claude Code daily, have opinions about how commits and reviews should be phrased, and would rather edit a markdown file than re-explain the same procedure in chat every morning.

Skills, agents, and how the two compose

Each skill is a directory containing a SKILL.md file. According to the README, that file holds YAML frontmatter with name, description and metadata, followed by instructions for Claude and optionally scripts, templates or references. Agents are a separate construct: an AGENT.md file with frontmatter that includes model selection, plus domain instructions and "decision-making frameworks." The README describes the relationship as bidirectional. Skills can decide when to invoke an agent, and agents can use skills while they work. In the repository layout, the engineering-workflow-plugin is the only plugin shown with an agents directory, and it contains exactly one agent, plan-implementer.md. That is worth noting before you install: the multi-agent orchestration implied by the vocabulary is, in this repository, a single planning agent plus four skills. The other three plugins are skills-only.

The execution runtime and its claimed token savings

The most consequential piece is execution-runtime, described as an implementation of the Anthropic code execution pattern. The mechanism is stated plainly: instead of loading code into context, Claude executes Python locally with API access. The repository splits it into an importable API library under api/ and a FastMCP server under mcp-server/, with setup.sh as a one-command installer. The README claims 90 to 99 percent token reduction for bulk operations, and gives one worked example: renaming getUserData to fetchUserData across Python files, quoted at roughly 25,000 tokens without execution and roughly 600 with it, a 97.6 percent saving. Those figures are the author's own and I have not reproduced them. Treat them as a hypothesis about your workload, not a specification. The runtime is also where the activation logic lives: the README says skills switch to execution mode for bulk operations of 10 or more files, complex multi-step workflows, iterative processing, and codebase-wide analysis. Below that threshold, the ordinary read-and-edit path applies, and the runtime buys you nothing.

Getting it installed, and the part that needs a second command

Installation is two separate steps, and the README does not merge them. First, inside Claude Code, you add the marketplace with /plugin marketplace add mhattingpete/claude-skills-marketplace, which the README says installs the engineering-workflow-plugin with all its skills and the plan-implementer agent. Individual plugins go in by path instead: /plugin marketplace add mhattingpete/claude-skills-marketplace/visual-documentation-plugin, and the same form for productivity-skills-plugin and code-operations-plugin. Second, and only if you want the execution runtime, you run the setup script directly from the installed plugin directory: ~/.claude/plugins/marketplaces/mhattingpete-claude-skills/execution-runtime/setup.sh. That path is a real constraint. It assumes the marketplace was installed under the name mhattingpete-claude-skills in that exact location. If your Claude Code installation uses a different plugins root, the command as printed will not resolve, and the README offers no alternate path or environment variable.

Where local execution is the wrong answer

Executing Python locally against your working tree changes the risk profile in a way that context-based editing does not. A model that reads a file and proposes an edit produces a diff you can reject. A script that runs with API access and writes across many files produces a result whose scope you have to reconstruct afterwards. The README lists PII and secret masking and sandboxed execution among the runtime's properties, but it does not enumerate what the sandbox restricts, and I cannot confirm from the supplied material whether writes are staged, reversible, or applied in place. For a repository with a clean git state that is a manageable question. For one without it, the execution path is the wrong tool. The same applies to small jobs. A three-file rename does not clear the 10-file activation threshold, so the runtime sits idle while you carry the setup and maintenance cost of a FastMCP server and a Python environment.

Compared with writing your own SKILL.md

The direct alternative is not another marketplace. It is a SKILL.md file you write yourself, dropped into your own Claude Code configuration. The difference is in who owns the trigger description. A hand-written skill encodes your conventions: your commit message format, your definition of a passing test, the files your reviewer should never touch. A marketplace skill encodes someone else's, and the README's own framing makes the trade-off explicit, since the model decides when to fire based on the description text. A vague description means the skill activates when you did not want it; a narrow one means it never fires. The marketplace version saves you the first draft and gives you a working structure to copy. It does not save you the editing, unless your conventions happen to match the ones shipped. The visual-documentation-plugin is the clearest case in either direction: flowchart, dashboard and timeline generation are generic enough that a shipped skill is plausibly close to what you would write, while git-pushing is exactly the kind of skill where your team's habits differ from everyone else's.

Licence, maintenance, and what the repository does not tell you

The repository is Apache-2.0, which permits commercial use and modification and includes an explicit patent grant, with the usual requirements around retaining notices and stating changes. That is a permissive baseline, but it governs the repository contents, not the Claude Code runtime or the Anthropic API you call through it, and it says nothing about what the execution runtime does with data it reads. No releases were retrieved for this repository, so there is no versioned artifact to pin against and no changelog to read before upgrading. Updating means pulling the current state of main. The plugins are markdown and small scripts, so the review cost per update is low, but the execution runtime is the exception: it is Python you install and run locally, and it will track whatever the FastMCP server and its API library require at the time you pull. Budget for reading execution-runtime/README.md on each update rather than assuming the setup script is idempotent.

Editorial conclusion

Adopt it if you already run Claude Code and want git, test-fixing and review behaviour packaged as model-invoked skills rather than hand-written slash commands; the Apache-2.0 licence and the plain SKILL.md layout make it cheap to read before you install. Skip it if your work is small, single-file edits, since the execution runtime only pays off on bulk operations and adds a local Python dependency you then have to maintain. Before installing, open execution-runtime/README.md and engineering-workflow-plugin/skills/git-pushing/SKILL.md and read what the skills actually instruct Claude to do, because the marketplace install command pulls in the whole engineering-workflow-plugin rather than one skill.

Official sources

1. Issues
2. License: Apache-2.0
3. mhattingpete/claude-skills-marketplace on GitHub
4. README