convez.ai
MCP ServerDados & AnalyticsIntermediário
thorstenfoltz

Octa - Explorador de Dados Local com MCP Server

thorstenfoltz/octa

Explorador de dados local rápido para 20+ formatos com MCP para IA

Ver no GitHub
InicialPluginScore · de 0 a 100

Ferramenta nativa em Rust para exploração interativa de 20+ formatos de dados (Parquet, DuckDB, SQLite, Jupyter, GeoJSON, etc). Oferece GUI com visualizações, queries SQL, exportação de gráficos, modo comparativo e funciona como MCP server para assistentes de IA. Inclui recursos como resumo estatístico de tabelas, busca com filtro/destaque e congelamento de colunas.

Instalação

$npx convez-ai thorstenfoltz/octa

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
36
Forks
1
Acompanhando
1
Saúde
42%
Último commit
há 5 dias
Versão
0.22.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Análise exploratória de dados em múltiplos formatos sem servidor
  • Assistência por IA em exploração de dados via MCP
  • Comparação e limpeza interativa de datasets
  • Geração de gráficos e exportação de análises

Veredito

Octa é uma solução robusta para exploração de dados local com integração MCP bem pensada. A GUI é produtiva, suporta muitos formatos nativamente e o recurso de MCP server permite workflows assistidos por IA sem sair da ferramenta. Ideal para analistas e engenheiros que lidam com datasets heterogêneos.

README

A tradução para o português deste README ainda está na fila.

What's new in this branch

This branch (feature/investigation) adds three user-facing features to Octa, plus a round of safety hardening and an Arrow/Parquet export fix.

Summary tab

A one-click overview of the active table: one row of statistics per column, the GUI counterpart of octa --describe and pandas' df.describe().

  • Open it from Analyse -> Summary...; it lands in a new tab named Summary - <file>.
  • The statistics come from DuckDB's SUMMARIZE, so each column gets its type, min / max, approximate distinct count, mean / standard deviation and quartiles (numeric columns), non-null count, and null percentage.
  • Unsaved cell edits are included: the summary describes the table as you currently see it, not the file on disk.
  • The result is an ordinary, detached table tab with no source path, so it can be sorted, filtered, copied, and exported via File -> Save As, but can never overwrite the original file. Re-run it after further edits for a fresh snapshot.

Search: filter or highlight

The toolbar search gains a mode toggle beside the search box that controls how matches are shown.

  • Filter (the default): non-matching rows are hidden, as before.
  • Highlight: every row stays visible and matching cells are highlighted in place.

Set the default under Settings -> Search & Editor -> Search result display; the toggle overrides it for the current session.

The table view honours the toggle. The text and tree views always highlight (hiding free text or collapsing tree nodes is not meaningful there), which means search now works in views where it previously had no effect:

  • Jupyter notebooks
  • JSON and YAML trees
  • Markdown (preview and editor)
  • the raw text editor
  • the in-app documentation dialog

Match count and navigation

When matches are highlighted, the search bar shows a current / total count and two buttons to step through them. While the search box is focused, Enter jumps to the next match and Shift+Enter to the previous one. The view scrolls the current match into view and emphasises it: in a table the match cell is selected, in text views the cursor moves to it, and in trees and notebooks the relevant node or cell scrolls into view.

Freeze columns

Pin key columns so they stay visible while scrolling a wide table horizontally, like freezing panes in a spreadsheet.

  • Right-click a column header and pick Freeze columns up to here to pin that column and every column to its left.
  • A thin separator marks the boundary; the rest of the table scrolls underneath. Unfreeze all columns in the same menu reverts.
  • The freeze is per tab and session-only, like column widths. If the window gets too narrow to keep the whole frozen band and still scroll, Octa temporarily pins fewer columns and restores the full band when there is room again.

Safety hardening

  • Assistant writes are confined to the export directory. The in-app chat assistant can only create files in the configured export folder (change it in Settings), with the single exception of writing back to a file you already have open in a tab. Writes anywhere else on disk are refused.
  • Tighter permissions on sensitive config and chat-transcript files.
  • Verified updates: release archives are checked against the published SHA256SUMS before the auto-update applies them.
  • Stricter archive extraction: tighter size limits and path checks when extracting entries from .zip / .tar / .tgz archives.

Arrow / Parquet export fix

Arrow and Parquet writes now cover every declared data type correctly, including width-correct numeric builders and proper binary and date handling, so schemas round-trip without mismatches.


📚 Full documentation: https://thorstenfoltz.github.io/octa/

Ver o original no GitHub

Bom para

Relacionados