Tapioca

tapioca

We bring the base. You pick the filling.

---

Distributable plugin for Claude Code and Cursor. tapioca is a neutral base — you add fillings (skills + companion agents) under one namespace.

PT — A nossa base, o seu recheio.

Plugin com skills e agents sob o namespace tapioca. A marca é a base; cada skill é um recheio. Algumas fillings são de domínio PT-BR (ex.: humanizer-br); outras são agnósticas (ex.: multi-gen, google-sheets-apps-script).

Included skills

/tapioca:multi-gen

Runs multiple AI CLIs in parallel (at least codex and cursor-agent) from an image briefing, validates each artifact (well-formed SVG), and builds a comparative HTML preview (light + dark + favicon strip at 16/32/48/180px) so you can pick a winner before manual polish.

• Portable parallel dispatch (macOS without gtimeout)
• Auth pre-check for cursor-agent (SKIP instead of hanging)
• Robust `` extraction when CLI stdout mixes logs
• Per-engine output: raw.txt, out., status.json in the run dir

/tapioca:multi-gen "geometric logo for Novarum, blue palette" --palette "#2E5BFF,#FFFFFF"

Pairs with a preview-server skill (control-plane) to serve index.html.

/tapioca:usabilidade-br

Audits web app usability against Jakob Nielsen's 10 heuristics. Captures evidence via Chrome MCP, optionally correlates with source (--code ), and writes a local self-contained HTML report with:

• Overall score (0–100) and per-heuristic scores
• Visual evidence (embedded screenshots)
• Code snippets with file:line when the component is found
• Copyable fix prompt per violation — paste into another Claude Code session

Companion agent parallelizes 10 passes (one per heuristic) and builds the report.

/tapioca:usabilidade-br http://localhost:3000 --code ./src

/tapioca:humanizer-br

Removes AI writing tells from Brazilian Portuguese and injects a human voice. Catalog of 25+ patterns: promotional tone, stacked gerunds, negative parallelism, rule of three, English-style Title Case, curly quotes, dramatic hooks, chatbot residue, and more.

Modes:

• Claude-only (default): runs in-session, zero extra cost.
• Optional Maritaca: delegates rewrite to sabia-3 (native PT-BR). Enabled when MARITACA_API_KEY is set.

Companion: humanizer-br agent (under agents/) for multi-pass with self-scoring.

/tapioca:google-sheets-apps-script

Google Sheets and Google Docs automation via Apps Script (browser mode — no GCP for the end user). Docs are markdown-first. The agent runs the CLI; the user pastes links and confirms in Google's UI.

/tapioca:google-sheets-apps-script

Upstream: gist v2.6.

Install

Claude Code

git clone https://github.com/gabriel-dantas98/tapioca ~/.claude/plugins/tapioca

Restart Claude Code or run /reload-plugins.

Cursor

git clone https://github.com/gabriel-dantas98/tapioca ~/.cursor/plugins/tapioca

The repo ships manifests for both platforms (.claude-plugin/plugin.json and .cursor-plugin/plugin.json) over the same skills/ and agents/.

Via --plugin-dir (Claude Code, local dev)

claude --plugin-dir /path/to/tapioca

Local smoke (marketplace + dual-CLI)

Harness in .agents/skills/smoke-test-skills/ — same source as CI:

.agents/skills/smoke-test-skills/run.sh marketplace                 # token-free
.agents/skills/smoke-test-skills/run.sh oneshot humanizer-br both   # uses tokens
.agents/skills/smoke-test-skills/run.sh all humanizer-br both

Cursor IDE: the harness creates ~/.cursor/plugins/local/tapioca (symlink). Then: Developer → Reload Window.

Usage

/tapioca:humanizer-br Paste the text to humanize here.

Or natural language:

Humanize this text, strip the AI smell.

For a serious multi-pass review, the companion agent is invoked automatically.

Maritaca mode

export MARITACA_API_KEY="your-key-here"

The skill detects the env var and offers Maritaca when the text justifies it (long enough, or quality explicitly requested).

Layout

tapioca/
├── .claude-plugin/plugin.json
├── .cursor-plugin/plugin.json
├── AGENTS.md
├── CLAUDE.md
├── README.md
├── skills/
│   ├── humanizer-br/
│   ├── usabilidade-br/
│   ├── multi-gen/
│   └── google-sheets-apps-script/
└── agents/
    ├── humanizer-br.md
    └── usabilidade-br.md

Roadmap

| Version | Content | |---|---| | v0.1 | humanizer-br (skill + agent), plugin format, AGENTS.md | | v0.2 | usabilidade-br, multi-gen | | v0.3 | google-sheets-apps-script, marketplace manifests | | next | More fillings; official marketplace submissions |

No hooks, MCP, or LSP in the short term. Skills and agents until the collection justifies more.

Credit

humanizer-br descends from:

• blader/humanizer (MIT) — canonical English humanizer for Claude Code / OpenCode.
• mackswendhell/humanizer-pt-br (MIT, 2026) — first direct WikiProject AI Cleanup adaptation to PT-BR.

What tapioca adds: namespaced plugin, Claude Code + Cursor, optional Maritaca sabia-3, multi-pass companion agents, and room for fillings beyond writing.

License

MIT — see LICENSE.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gabriel-dantas98
• Source: gabriel-dantas98/tapioca
• License: MIT
• Homepage: https://gabriel-dantas98.github.io/tapioca

Install and usage instructions live in the source repository linked above.