convez.ai
MCP ServerDevOps & InfraestruturaAvançado
Ethirajulu

Agentic Debugger

Ethirajulu/agentic-debugger

Investigação automática de bugs em SaaS com múltiplas fontes de dados via MCP

Ver no GitHub
InicialPluginScore · de 0 a 100

Sistema de depuração agentic para triagem de issues de clientes em produtos SaaS. Integra múltiplas fontes de dados (logs, codebase, contexto de cliente) via constelação de servidores MCP e aplica raciocínio agentic (ReAct) para investigar automaticamente e produzir relatórios acionáveis.

Instalação

$npx convez-ai Ethirajulu/agentic-debugger

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
28%
Último commit
há 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Triagem automática de issues reportadas por clientes
  • Investigação multi-repo de bugs complexos
  • Geração de relatórios de triage com causa raiz e recomendações

Veredito

Projeto educacional de engenharia de IA com foco em padrões agnósticos (MCP constellation + ReAct). Prático para times SaaS que enfrentam triage manual de bugs, mas ainda em fase de curriculum — a implementação completa segue cronograma de 5 fins de semana.

README

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

Agentic Debugger

An AI debugging assistant for customer-reported issues in SaaS products. Investigates issues by reading logs, searching across multiple code repos, fetching customer context, forming hypotheses, and producing an actionable report — the way a senior on-call engineer would, but automated.

Learning project. Built to attack a real pain (multi-repo, multi-source bug triage) using an architecturally interesting pattern (MCP server constellation + agentic client). Follows on from my previous AI engineering project — mini_customer_support_agent — which covered the foundations (RAG, agents, eval, observability).

What it does (eventually)

Customer issue   →   [agent investigates]
                         ├─ logs-mcp: find error spikes, traces, stack traces
                         ├─ codebase-mcp: search across many repos, read files, blame lines
                         ├─ customer-mcp: tenant info, plan, version, prior tickets
                         └─ hypothesize → verify → repeat
                     ↓
                 Investigation report
                  • Root cause
                  • Evidence (log lines, code refs, customer data)
                  • Recommended fix or escalation

Why this shape (MCP constellation, not monolithic agent)

The Model Context Protocol lets you expose data sources as independent servers that any compatible client (Claude Desktop, Cursor, custom UIs) can talk to. Three benefits over a monolithic app:

  • Each server is a vertical slice — logs-mcp is useful by itself in Claude Desktop, before any agent exists.
  • Composable across teams — other engineers can build MCP servers for their domains and the same agent uses all of them.
  • The agent stays thin — it's a ReAct loop over a growing toolset, not a custom integration per data source.

Curriculum

This project is a five-weekend build, plus extensions. Detailed plan in docs/curriculum.md (coming).

WkBuildWhat it teaches
1logs-mcp server (Datadog/Honeycomb)MCP protocol basics, server design
2codebase-mcp server (multi-repo semantic + lexical search)Code RAG, pgvector revisited
3customer-mcp server (mock or real customer data)Domain modeling at the tool boundary
4Agent client — ReAct loop over all three MCP serversMulti-source synthesis, the killer demo
5Eval + iteration on a golden set of past customer issuesClosing the measurement loop

Extensions (drawn from advanced agentic patterns):

  • Refactor to multi-agent (triage agent → specialist agents per data source)
  • Add durable workflows (investigations span days, pause for human approval)
  • Add persistent memory (agent remembers prior bug patterns and applies them)

Stack

  • Node.js 22 + TypeScript
  • Anthropic SDK (Claude Haiku for agent, Sonnet for judge — same defaults as the predecessor project)
  • Ollama support for fully-local runs (llama3.1:8b chat, qwen2.5:14b judge)
  • Postgres 16 + pgvector for code embeddings
  • LangFuse for tracing every LLM call + judge scores
  • MCP TypeScript SDK for protocol implementation

Quick start (placeholder — fills in as weeks land)

pnpm install
cp .env.example .env.local
# Add ANTHROPIC_API_KEY at minimum
docker compose up -d           # Postgres + pgvector on port 5433
pnpm typecheck

Week-specific run commands will be added per phase.

Project layout (current — grows over weeks)

agentic-debugger/
├── lib/
│   ├── llm.ts                  router (anthropic | ollama)
│   ├── llm-anthropic.ts        Anthropic chat wrapper
│   ├── llm-ollama.ts           Ollama chat wrapper
│   ├── judge.ts                LLM-as-judge for debugging-domain rubrics
│   └── tracing.ts              LangFuse SDK wrapper (no-ops if not configured)
├── docs/                       curriculum doc lives here
├── scripts/                    eval runner, indexing scripts (per week)
├── docker-compose.yml          local Postgres + pgvector
└── .env.example

Subdirectories per MCP server will land in week order: mcp-servers/logs/, mcp-servers/codebase/, etc.

Predecessor

The foundations this project reuses — RAG retrieval, agent loop patterns, eval pipeline, LangFuse instrumentation — were built and documented in mini_customer_support_agent. Blog series at kumizhi-ai.com/blog.

Ver o original no GitHub

Bom para

Relacionados