convez.ai
MCP ServerDevOps & InfraestruturaIntermediário
frontdoorrr

LogSqueeze

frontdoorrr/logsqueeze

Compressão de logs drop-in para Claude Code e agentes MCP

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que comprime logs usando o algoritmo XDrain, reduzindo milhões de linhas em templates compactos. Integra-se nativamente com Claude Code e agentes MCP, processando logs localmente sem enviar dados brutos para o modelo.

Instalação

$npx convez-ai frontdoorrr/logsqueeze

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
0
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 5 meses
Versão
v0.1.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Analisar logs grandes sem sobrecarregar o contexto do modelo
  • Debugar aplicações via Claude Code com históricos volumosos
  • Processar saída de comandos (kubectl, docker, journalctl) em tempo real

Veredito

Ferramenta focada e bem executada para um problema real em workflows de IA: logs volumosos esgotam contexto rápido. A compressão local via XDrain é transparente e eficaz, com suporte a múltiplos formatos. Recomendado para equipes que usam Claude Code em debug/ops.

README

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

English | 한국어

logsqueeze

CI Go Reference License: MIT

Drop-in log compression for Claude Code and any MCP-compatible agent.


Why

Claude Code's context fills up fast when you paste large log files. logsqueeze compresses logs locally using the XDrain algorithm — collapsing millions of lines into a handful of templates — before they ever reach the model. No data leaves your machine.


How it looks

Compressed 1,200,000 lines → 12 templates (100,000x compression)

[x1,185,000] worker ready shard=<*> [1..48]
  samples: 14:22:11 worker ready shard=1 | 14:22:12 worker ready shard=2

[x12,000] pool acquire <*> [20..480ms p50=240ms]
  samples: 14:22:15 pool acquire 240ms | 14:22:16 pool acquire 480ms

[x3] ERROR psycopg2.OperationalError: connection <*> [timeout,refused,reset]
  samples: 14:22:16 ERROR psycopg2.OperationalError: connection timeout

Each <*> slot summarizes what varied in that position: numeric ranges with p50, or a list of distinct string values.


Install

go install github.com/frontdoorrr/logsqueeze@latest

Claude Code setup

Add to ~/.claude.json:

{
  "mcpServers": {
    "logsqueeze": {
      "command": "logsqueeze",
      "args": ["mcp", "serve"]
    }
  }
}

Restart Claude Code and the three tools below are available in every session.


MCP Tools

ToolDescription
compress_logsCompress raw log text passed inline
compress_fileRead a log file by path and compress it
compress_commandRun a shell command and compress its stdout

Examples Claude Code can call:

# Inline log text
compress_logs(logs="<paste log content>")

# Log file
compress_file(path="/var/log/app.log")

# Live command output
compress_command(command="kubectl logs -n prod deploy/api --tail=5000")
compress_command(command="docker logs --tail=2000 my-container")
compress_command(command="journalctl -u nginx --since '1 hour ago'")

How it works

logsqueeze implements XDrain, a log template mining algorithm that groups lines by structural similarity using a prefix tree. Incoming lines are processed in shuffled batches (removing ordering bias), and each line is tried against multiple token rotations before a group is chosen by majority vote. Variable positions become <*> wildcards; numeric slots accumulate all seen values for accurate min/max/p50 statistics; string slots collect up to 6 distinct samples.

The result is a compact, human-readable summary that gives an LLM the statistical shape of the logs without the raw volume.


Supported log formats

logsqueeze parses automatically — no format flag required in most cases.

FormatExample
ISO 8601 syslog2024-01-01T14:22:11Z INFO worker ready
Space-separated syslog2024-01-01 14:22:11 [INFO] worker ready
Slash-date syslog2024/01/01 14:22:11 INFO worker ready
JSON (structured){"time":"...","level":"info","msg":"worker ready"}
Level-onlyINFO: worker ready / [ERROR] pool acquire 240ms
Plain textworker ready shard=1

JSON fields recognized: time/timestamp/ts/@timestamp, level/severity/lvl, msg/message/log/text.


License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados