mstack

Claude Code plugin marketplace for development and continuous-improvement workflows.

Install

/plugin marketplace add mayank-io/mstack

Then enable individual plugins:

/plugin install notes@mstack
/plugin install fetch@mstack
/plugin install dev@mstack
/plugin install ccimprove@mstack
/plugin install company-setup@mstack
/plugin install think@mstack

notes and fetch are a pair — notes:clip routes to fetch:* for everything it captures, so installing notes alone leaves most routes dead.

Plugins

ccimprove — meta improvement

Analyze your Claude Code usage and turn one-off work into reusable patterns.

Skill	What it does
/ccimprove:surface-usage-patterns	Scan all sessions in ~/.claude/projects/ and surface candidates for skills, plugins, agents, and CLAUDE.md additions
/ccimprove:make-repeatable	Analyze the current conversation and recommend whether to codify it as a skill or a plugin, then build it
/ccimprove:clean-permissions	Prune one-off pastes, redundant entries, and dead MCP refs from ~/.claude/settings.json

Ships a hook that runs automatically: claude plugin update leaves the old commit-sha directory in the cache, and that directory stays loaded — every stale copy contributes its skills to the slash menu again. Three copies of /fetch:x-post appeared this way. The hook prunes any cache directory no longer referenced by installed_plugins.json, on PostToolUse/Bash when the update runs through Claude and on SessionStart (throttled hourly) for updates made via /plugin or a plain terminal. Run it by hand with python3 plugins/ccimprove/scripts/prune_plugin_cache.py --dry-run.

dev — development workflow

Iterative reviews, doc conventions, and feedback loops for serious engineering work.

Skill	What it does
/dev:code-review-iterative	Up to 4 review-fix iterations with 3 parallel reviewer agents (PE, Sr SDE, QA), diminishing fix thresholds
/dev:design-review-iterative	Same shape, applied to design documents (PE, Sr SDE, Domain Expert)
/dev:ml-design-review-iterative	ML-specific variant (Principal ML Engineer, Sr Applied Scientist, ML Ops Engineer)
/dev:apply-feedback	Apply tagged comments ([MK], [REV], etc.) in the current document immediately, no questions asked
/dev:review-feedback	Same input, but plan-then-apply with approval gate
/dev:setup-doc-conventions	Bootstrap docs/templates/, dashboard, Obsidian config, and CLAUDE.md workflow section
/dev:finish-work-on-local-worktree	Squash-merge a worktree into main, update doc statuses, refresh the project plan, recommend next work
/dev:summarize-this-session	Two-section conversation recap: ongoing task + continuous-improvement insights

Bundled assets:

• 10 doc-type templates (design, impl-plan, cutover-plan, analysis, report, postmortem, kb, ops-guide, playbook, experiment)
• Obsidian Bases dashboard (dashboard.base + dashboard.md)
• CLAUDE.md workflow section snippet
• PreToolUse / PostToolUse / SessionStart hooks for doc workflow enforcement

Configuration: feedback tags

apply-feedback and review-feedback look for tagged comments (e.g., [MK], [REV]) in the current document. The set of tags and how each one should be interpreted is configured per user at:

~/.mstack/dev/feedback-tags.json

The file is auto-created on first run with a default MK entry. To add, modify, or delete tags, edit this file directly. Changes are picked up on the next skill invocation — no restart required.

Schema:

{
  "tags": [
    {
      "tag": "MK",
      "from": "Document author (initials = MK for Mayank)",
      "content": "Direct edit instructions, factual corrections, rephrasing requests",
      "action": "Apply edits as written; treat as imperative; only push back if ambiguous"
    },
    {
      "tag": "REV",
      "from": "External reviewer",
      "content": "Suggestions, open questions, alternative phrasings",
      "action": "Treat as suggestions; flag for discussion before applying"
    }
  ]
}

Field semantics:

• tag — short uppercase prefix that appears bracketed in documents (matched case-insensitively).
• from / content — context to help Claude calibrate trust and tone for these comments.
• action — interpretation guidance. Does not override the skill's posture (apply-feedback still applies immediately, review-feedback still gates on approval), but informs how matched comments are treated within that posture.

notes — capture a source into your vault

Start here. clip is the single entry point: hand it a URL and it routes to the right fetcher, then files the result into whichever Obsidian vault you are running in.

Skill	What it does
notes:clip	URL in, note out — routes by host, chains on the fetcher's result line
notes:create	Write a note into the current vault; owns frontmatter, folder, filename, collisions
notes:save-local-file	File already on disk → note, archiving the file into the vault's attachments

The vault is wherever the session is — found by walking up for a .obsidian/ directory. There is no vault_path config.

fetch — pull content from a source into a directory

The layer beneath notes. Each skill gets content out of one source and knows nothing about vaults. Useful directly when you want the raw material somewhere of your choosing.

Skill	What it does
fetch:youtube-transcript	Transcript + metadata + chapters + speakers from a YouTube URL (persistent Chrome profile, Whisper fallback, caption-integrity verification)
fetch:x-post	Single tweet, full thread, or X Article — auto-detects threads, walks back to the thread root, downloads images at original resolution
fetch:linkedin-post	Post text, author, metrics, comments and attachments — expands the "… more" control and filters attachments from page furniture
fetch:blog-post	Article to self-contained Markdown + images, recovering the lazy-loaded ones Defuddle drops
fetch:notion-public-site	Crawl a public Notion site, save every page as Markdown with wikilinks and embedded images
fetch:scribd-document	Pull every page of a Scribd document as zero-padded .jpg files via embed view
fetch:alphaxiv-paper	Structured AI-generated overview of any arXiv paper from alphaxiv.org
fetch:vedic-chart	Digitize a Vedic astrology chart image into structured data

Contract. Every fetch:* skill takes <url-or-path> [output_dir] — a temp directory when omitted — never writes outside it, and prints OUTPUT_FILE:<path> or OUTPUT_DIR:<path> as its final stdout line. Chain on that line; never reconstruct the path.

Conventions

• Skills only — no command wrappers. A skill is both slash-invocable (/notes:clip) and naturally triggered by its description, so it needs no companion command. Until 0.9.0 every skill had a same-named wrapper in commands/ whose entire body was "Invoke the <plugin>:<name> skill via the Skill tool." Both registered the same /plugin:name, so all 23 appeared twice in the slash menu — and the two descriptions drifted apart. Removing them cut the menu from 74 entries to 51 and, because the skill's own description is longer than the wrapper's, made the remaining entries more informative rather than less.
• No Claude attribution in commits. See ~/.claude/CLAUDE.md.
• Docs live elsewhere. Design docs and implementation plans for this marketplace live in the sibling repo mstack-docs/, not in this repo.

Layout

mstack/
├── .claude-plugin/
│   └── marketplace.json    # marketplace manifest
└── plugins/
    ├── ccimprove/
    ├── company-setup/
    ├── dev/
    ├── fetch/
    ├── notes/
    └── think/

Each plugin follows the standard layout:

<plugin>/
├── .claude-plugin/plugin.json
├── skills/<name>/SKILL.md  # implementations
├── hooks/                  # (dev only)
├── scripts/                # (dev, fetch)
└── templates/              # (dev only)

License

MIT