Reka — Plugin Claude Code
getreka/reka-plugin
RAG para IA development: busca semântica, memória de projeto e workflows
Ver no GitHubMCP Server que conecta Claude a uma infraestrutura RAG compartilhada, fornecendo busca semântica no codebase, memória persistente de projeto, análise de arquitetura e workflows estruturados de desenvolvimento. Mantém contexto e decisões entre sessões de múltiplos usuários.
Instalação
npx convez-ai getreka/reka-plugin1 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
Casos de uso
- Busca semântica em grandes codebases durante pair programming com Claude
- Manutenção de memória de decisões arquiteturais e padrões técnicos
- Análise de impacto antes de refatorações (blast radius)
- Code review estruturado com contexto histórico do projeto
Veredito

Solução sofisticada para times que usam Claude Code em projetos complexos, com foco em persistência contextual e workflows estruturados. Requer setup de RAG API externa, o que adiciona complexidade operacional mas retorna em qualidade das assistências de desenvolvimento.
README
A tradução para o português deste README ainda está na fila.
Overview
Reka is a Claude Code plugin that connects your AI assistant to a shared RAG (Retrieval-Augmented Generation) infrastructure. It gives Claude persistent project memory, semantic codebase search, architecture awareness, and structured development workflows — across sessions and team members.
What you get
| Capability | How it works |
|---|---|
| Semantic search | Search code, docs, and Confluence by meaning, not just keywords |
| Project memory | Decisions, patterns, and insights persist across sessions via Ebbinghaus-inspired retention |
| Architecture awareness | ADRs, dependency graphs, blast radius analysis before every change |
| Structured workflows | 5-phase coding, deep investigation, tribunal debates, and more |
| Auto session lifecycle | Hooks start/end RAG sessions automatically, trigger memory consolidation |
Quick Start
1. Add marketplace
/plugin marketplace add getreka/reka-plugin
2. Install plugin
/plugin install reka@reka-plugins
3. Configure
On first enable, Claude Code prompts for three settings:
| Setting | Description | Example |
|---|---|---|
| RAG API URL | Your RAG API server | http://localhost:3100 |
| RAG API Key | Authentication key (stored in OS keychain) | e699194c-... |
| Project Name | Collection namespace | myapp |
Team auto-install
Add to your project's .claude/settings.json so everyone gets it automatically:
{
"extraKnownMarketplaces": {
"reka-plugins": {
"source": { "source": "github", "repo": "getreka/reka-plugin" }
}
},
"enabledPlugins": {
"reka@reka-plugins": true
}
}
Commands
Coding & Review
| Command | Description |
|---|---|
/reka:code | 5-phase workflow: context → plan → implement → verify → remember |
/reka:investigate | Deep research — find, trace, debug. Saves to memory, never modifies code |
/reka:review | Architecture-aware code review against patterns and ADRs |
Architecture
| Command | Description |
|---|---|
/reka:arch | Record and analyze architecture decisions (ADRs) |
/reka:debate | Adversarial tribunal debate for complex decisions (2-4 positions) |
Session & Memory
| Command | Description |
|---|---|
/reka:start | Display session status and project stats (sessions auto-start via hook) |
/reka:end | Save knowledge, close session, trigger memory consolidation |
/reka:memory-review | Triage quarantine queue, promote/reject auto-extracted memories |
Setup
| Command | Description |
|---|---|
/reka:onboard | Set up RAG for a new project: configure, index, verify |
Agents
Internal subagents used by the workflows — /reka:code, /reka:investigate, and /reka:review delegate to them via the Task tool. For direct requests, prefer the corresponding /reka:* command.
| Agent | Model | Used by |
|---|---|---|
reka:feature-builder | Sonnet | /reka:code Phase 3 — multi-file implementations with RAG context |
reka:code-reviewer | Sonnet | /reka:review — per-file review against patterns and ADRs |
reka:rag-researcher | Haiku | /reka:investigate — parallel semantic search and graph traversal |
reka:rag-ops | Haiku | Operations: indexing, collections, memory maintenance |
All agents have persistent memory (memory: project) — they learn your codebase patterns across sessions.
Hooks
| Event | Action |
|---|---|
| SessionStart | Auto-starts RAG session, injects RAG_SESSION_ID env var, injects the session digest into context |
| SessionEnd | Ships the session transcript to your RAG API, ends the RAG session, triggers consolidation agent |
Transcript capture sends the last 2 MiB of the Claude Code transcript to POST /api/capture/transcript on the RAG API URL you configured — your own server is the only destination; the plugin sends data nowhere else. Set REKA_TRANSCRIPT_CAPTURE=0 to disable it.
Skills
Reference skills loaded automatically by commands and agents:
- memory-protocol — Session lifecycle, smart remember with relationship detection, memory type selection, structured facts
- rag-workflows — Search tool priority guide (Grep → find_symbol → hybrid_search → search_graph → context_briefing)
- obsidian-sync — Bidirectional sync between RAG memories and Obsidian vault
Architecture
Claude Code
│
└── Reka Plugin
├── 9 Commands (/reka:code, /reka:investigate, ...)
├── 4 Agents (feature-builder, code-reviewer, ...)
├── 3 Skills (memory-protocol, rag-workflows, obsidian-sync)
├── 2 Hooks (session lifecycle)
│
└── MCP Server (@getreka/mcp)
│ 29 tools: search, memory, architecture, sessions, agents
│
└── RAG API
├── Qdrant — vector database
├── Ollama — embeddings (qwen3-embedding) + local LLM
└── Claude — complex tasks (optional, hybrid routing)
Prerequisites
- RAG API server running (getreka/reka)
- Qdrant vector database
- Ollama — embeddings + local LLM (default provider)
Optional but recommended:
typescript-lsp@claude-plugins-officialfor TypeScript code intelligence- Obsidian for memory visualization (via
/reka:obsidian-syncskill)
Local Development
claude --plugin-dir ./reka-plugin
Use /reload-plugins after making changes to pick up updates without restarting.
Test the hooks (plain bash, mocks curl via a PATH shim):
bash tests/session-start.test.sh
bash tests/session-end.test.sh