Mewbo – Orquestrador de Agentes IA
bearlike/Assistant
Spawn, escopo e coordena agentes IA paralelos com grafo de conhecimento integrado
Ver no GitHubStack aberto para automação de trabalho com agentes IA paralelos, com gerenciamento automático de sub-agentes e documentação em grafo de conhecimento. Funciona em terminal, navegador, API ou Home Assistant, aceitando qualquer modelo de linguagem.
Instalação
npx convez-ai bearlike/Assistant1 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
- Automação de tarefas complexas que exigem agentes paralelos
- Documentação e Q&A baseados em grafo de conhecimento
- Busca sintética agregada em múltiplas ferramentas
- Orquestração de fluxos agenticos com aprovação de planos
Veredito

Ferramenta robusta e bem arquitetada para cenários exigentes de orquestração de agentes. Requer familiaridade com sistemas distribuídos, mas oferece controle fino e documentação técnica sólida.
README
A tradução para o português deste README ainda está na fila.
https://github.com/user-attachments/assets/78754e8f-828a-4c54-9e97-29cbeacbc3bc
Overview
Mewbo is an open, model-agnostic stack for agentic work. At its core, a hypervisor decomposes a goal into parallel sub-agents that each carry only the tools they need and exchange compressed summaries instead of raw transcripts, within resource budgets you tune per deployment. Three products build on that core. The first automates multi-step tasks across your codebase and tools, isolating each change in its own Git worktree. The second generates living documentation and question answering, grounded in a multiplexed memory graph that runs alongside an AST index of your code. The third is a search engine rebuilt around agents and indexes, ranking results across every source you have connected. You approve destructive actions, watch the agent tree as it grows, and can steer any branch mid-flight. Sessions persist with full provenance, compact automatically near the budget, and run on any model across every client.
Features
- Agent hypervisor. Sub-agents spawn in parallel with scoped tools and approval-gated actions. Progress shows as a live tree, and you can steer or cancel any branch mid-flight. The hypervisor enforces resource budgets through natural-language warnings rather than force-kills, and resolves every child into a structured result.
- Long-horizon context. Two-mode compaction summarises older turns near the budget. Post-compact file restoration replays the working set. Conversation fork lets you branch from any message and replay against a different model.
- Native skills, plugins, and MCP. Agent Skills, plugins from any compatible marketplace, and MCP servers load from user or project scope without translation. Plugins also contribute per-session stateful tools, hooks, and agent definitions.
- Agentic Wiki. Turns any repository into living documentation you can interrogate. Indexing pairs a multiplexed memory graph with an AST index of the code, so question answering traverses structure and meaning across many hops for authoritative answers grounded in the source itself.
- Agentic Search. A search engine rebuilt around agents and indexes. One query fans out across every source you've connected, from repos to trackers to chat, and comes back as one ranked list of results spanning them all, topped by a synthesised overview cited to its sources.
- Inline interactive widgets. Sub-agents author Streamlit-in-WASM widgets that mount in a sandboxed Web Worker inside the conversation, with no server round-trip and no CORS.
- Provider-agnostic, multi-surface. Any model behind LiteLLM, accessed from a terminal CLI, web console, REST API, MCP server, Home Assistant, Nextcloud Talk, or email. Same session, same tools, same transcript.
Get started
See docs.mewbo.com/latest/getting-started to install Mewbo and run a first session.
Documentation
Full documentation lives at docs.mewbo.com.
| Section | Covers |
|---|---|
| Get Started | Install, configure an LLM, run a first session. |
| Configure | LLM setup, project config, configuration reference. |
| Clients | CLI, web console, REST API, Home Assistant, Nextcloud Talk, email. |
| Knowledge & Discovery | Agentic Wiki and Agentic Search, the source-grounded docs and cross-tool search products. |
| Capabilities | Built-in tools, sub-agents, skills, plugins, widgets, plan mode, permissions, compaction. |
| Deploy | Docker Compose, storage backends, production setup. |
| Develop | Architecture, session runtime, building a client, API reference. |
| Releases | Release notes and upgrade history. |
Contributing
Bugs and feature requests on the issue tracker. For development setup, see the developer guide.
License
MIT © Krishnakanth Alagiri.