convez.ai
MCP ServerDevOps & InfraestruturaAvançado
ninthhousestudios

Manas — Sistema de Cognição Colaborativa para Agentes IA

ninthhousestudios/manas

Sistema nervoso modular para agentes IA com memória, percepção e planejamento

Ver no GitHub
InicialPluginScore · de 0 a 100

Infraestrutura modular em Rust que fornece memória persistente, percepção de filesystem, coordenação, planejamento e inteligência de código para workflows com LLMs. Implementa subsistemas especializados (chitta, smriti, sutra, sangha, kosha, yojana, vidya) que rodam como daemons locais com servidor MCP, criando um "corpo" robusto para agentes de IA.

Instalação

$npx convez-ai ninthhousestudios/manas

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
1
Forks
0
Acompanhando
1
Saúde
42%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Workflows de desenvolvimento com agents Claude Code/Codex com estado persistente
  • Análise de inteligência de código e dependências em projetos grandes
  • Memória e raciocínio persistentes para assistentes pessoais IA
  • Coordenação de múltiplos agentes e tarefas locais

Veredito

Projeto ambicioso e bem arquitetado que trata IA local como cidadão de primeira classe. Exige experiência em Rust e Linux; ideal para quem quer construir agentes autônomos robustos com memória genuína, não simulada. Maduro o suficiente para exploração séria, ainda em desenvolvimento ativo.

README

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

manas

License: AGPL v3

manas (मनस् — "mind") is a collaborative cognition system for humans and agents. It's a federation of agent-native subsystems that provide memory, perception, coordination, planning, code intelligence, and document comprehension to LLM-driven workflows. The LLM is the cognition; manas is its body.

Each subsystem is a standalone Rust binary with its own MCP server surface, systemd unit, and README.

Subsystems

SubsystemSanskritWhat it doesStatus
chittaचित्त — consciousnessPersistent working model of a person: observations, decisions, mental models, session episodes. Postgres + pgvector, BGE-M3 ONNX embeddings, hybrid retrieval.v0.3.2
smritiस्मृति — remembranceContent-addressed filesystem indexer. Tracks files by BLAKE3 hash, detects moves/copies/edits, emits lifecycle events. SQLite + FTS5.v0.3.5
sutraसूत्र — threadCode intelligence via tree-sitter. Symbols, calls, deps, blast radius, hotspots, dead code, co-change analysis. SQLite.v0.7.8
sanghaसंघ — assemblySession coordination daemon. Registry with heartbeat TTL, advisory resource locks, broadcast inbox. SQLite.v0.1.0
koshaकोश — treasuryDocument comprehension. PDFs, EPUBs, scanned pages via multimodal embeddings (Qwen3-VL). Postgres + pgvector.v0.1.0
vidyaविद्या — knowledgeStructured knowledge graph with provenance. Cited, tradition-aware domain facts. Postgres.v0.4.1
yojanaयोजना — planLocal task graph. Projects, tasks, dependency edges, context shapes. SQLite.v0.3.3
manas-cli—Ops CLI: manas health, manas warm, manas done, manas serve. Harness adapters for Claude Code, Codex, Gemini.v0.1.2
vidhiविधि — methodEngineering methodology skills for Claude Code agents. Planning, TDD, diagnosis, architecture.v0.1.0
karmaकर्म — actionPersonal assistant layer: triggers, connectors, approval gate.design only

Architecture

Each subsystem runs as a local daemon (systemd user service) exposing MCP over streamable HTTP. The CLI (manas) ties them together for lifecycle operations — health checks, session warm-up, shutdown.

 Agent (Claude Code / Codex / Gemini)
   │
   ├── chitta  :3100  (memory)
   ├── smriti  :3300  (filesystem)
   ├── sutra   :3400  (code intelligence)
   ├── sangha  :3200  (coordination)
   ├── kosha   :3500  (documents)
   ├── yojana  :4200  (tasks)
   └── manas   :3000  (composed ops)

See docs/manas-architecture.md for the full design.

Install

Each subsystem builds independently:

cd <subsystem>
cargo install --path .

Most subsystems need a systemd user service. Check each README for prerequisites and setup.

Context

This is a personal project — built for my own AI workflow. It's public because I believe in open source and because the ideas might be useful to others building similar systems. It is not designed for general-purpose deployment and there is no support commitment. I have "stolen" many ideas from various sources; not source if they are mentioned but I will add that in the future if not. So while I want everyone to respect the license, the ideas and design choices I have made are free to inspect and build upon if you want.

License

AGPL-3.0 for all subsystems except vidhi (MIT, forked from mattpocock/claude-code-skills).

Ver o original no GitHub

Bom para

Relacionados