convez.ai
MCP ServerAutomaçãoAvançado
axocoatl

Axocoatl

axocoatl/axocoatl

Runtime persistent para agentes IA autocoordinados, 100% local

Ver no GitHub
Muito bomPluginScore · de 0 a 100

Runtime Rust para sistemas multi-agentes autocoordinados com persistência, checkpointing e memória de 4 camadas. Agentes ativam via lattice de eventos estigmérgico (sem orquestrador central), suportam múltiplos provedores e ferramentas MCP. Totalmente local, zero telemetria.

Instalação

$npx convez-ai axocoatl/axocoatl

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
74
Forks
11
Acompanhando
1
Saúde
100%
Último commit
há 2 semanas
Versão
v1.5.3
Issues abertas
1
Contribuidores
2

Casos de uso

  • Orquestração de agentes sem servidor central
  • Fluxos de trabalho tolerantes a falhas com checkpointing
  • Pesquisa multi-agente com coordenação automática
  • Sistemas agentic híbridos com LLMs locais e remotos

Veredito

Ferramenta sofisticada e diferenciada para quem quer arquitetura multi-agente descentralizada com persistência real. Implementação em Rust garante performance e confiabilidade, mas exige familiaridade com conceitos avançados de coordenação agentic e configuração YAML.

README

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

Axocoatl

The Rust runtime for self-coordinating multi-agent systems.

CI crates.io License

Axocoatl runs persistent AI agents that coordinate through a stigmergic event lattice — agents activate when their dependencies complete, driven by pheromone-style signals with no central orchestrator. Built in Rust on the ractor actor model: low memory, fast cold start, provider-agnostic.


60-second quickstart

# 1. Install (no Rust toolchain required)
curl -fsSL https://raw.githubusercontent.com/axocoatl/axocoatl/main/scripts/install.sh | sh

# 2. Interactive setup wizard — picks a provider, scaffolds a project
axocoatl onboard

# 3. Check your environment
axocoatl doctor

# 4. Start the daemon + API, then chat
axocoatl dev
axocoatl chat -a assistant

Prefer Cargo? cargo install axocoatl-cli (requires Rust 1.82+).

Skipping onboard? Copy axocoatl.example.yaml to axocoatl.yaml — two agents and one workflow, fits on one screen. The full axocoatl.yaml shipped in the repo is the larger demo (12 agents, scheduled runs, MCP servers).


Why Axocoatl

CapabilityAxocoatlAutoAgentsCrewAI
Language / runtimeRust / actorsRust / actorsPython
Stigmergic coordination (no orchestrator)✅❌❌
HTN symbolic planning✅❌❌
Auction-based agent selection✅❌❌
Per-agent token budgets✅❌partial
4-tier persistent memory + checkpointing✅partialpartial
MCP client + server✅partial✅
A2A protocol✅❌❌
Provider-agnostic (Ollama/OpenAI/Anthropic/…)✅✅✅
Interactive onboarding + doctor✅❌❌

The differentiator is the coordination layer: define agents with depends_on, and the event lattice cascades work through them automatically.

agents:
  - id: researcher
    provider: ollama
    model: llama3.2
    depends_on: []
  - id: summarizer
    provider: ollama
    model: llama3.2
    depends_on: [researcher]   # activates when researcher completes

workflows:
  - id: research-and-summarize
    agents: [researcher, summarizer]
    entry_point: researcher
axocoatl workflow run research-and-summarize -i "What is photosynthesis?"

Core concepts

  • Agents — persistent ractor actors with a provider, tools, 4-tier memory, and a token budget. Survive restarts via checkpointing.
  • Hybrid memory recall — relevant past exchanges are injected each turn, and the agent can also pull on demand: recall_search (semantic search over past sessions) and recall_timeframe (read a day's activity log). Tunable per agent.
  • Agent-managed core memory — editable blocks (persona, human, project, …) the agent curates via tools and that render into its prompt each turn (the MemGPT/Letta model). Per-agent by default, shareable across agents. A background "sleep-time" pass consolidates idle agents' memory automatically.
  • Stigmergic coordination — agents publish TaskCompleted events; an EventLattice accumulates pheromone signals and activates downstream agents when thresholds are crossed. No scheduler, no glue code.
  • Coordinator role — for explicit hierarchical work, an agent with role: coordinator decomposes a goal into subtasks (HTN or LLM), auctions them to worker agents, runs them in parallel, and synthesizes the results. The pass is resumable via checkpointing.
  • Workflows — declarative multi-agent DAGs via depends_on / entry_point.
  • Providers — Ollama, OpenAI, Anthropic, Mistral, Gemini, OpenRouter. No lock-in.
  • Protocols — MCP (consume & expose tools) and A2A (agent interop).

See the docs site for the full picture, the marketing site for the positioning, or docs/ARCHITECTURE.md and docs/TROUBLESHOOTING.md for the in-repo quick reference.


Roadmap

  • Stronger sandbox isolation tiers — the shipped sandbox is a hardened rootless Podman container (capabilities dropped, no-new-privileges, network-isolatable); microVM-class isolation (Firecracker) is planned.

CLI

axocoatl onboard                 Interactive setup wizard
axocoatl doctor                  Environment / dependency health check
axocoatl init <name>             Scaffold a project non-interactively
axocoatl validate <config>       Validate a config file
axocoatl dev | serve             Run daemon (+ IPC) / production server
axocoatl chat -a <agent>         Interactive chat
axocoatl workflow list | run     Inspect / execute multi-agent workflows
axocoatl agents list|status|restart
axocoatl tokens report           Per-agent token usage
axocoatl mcp servers|tools       Inspect connected MCP servers/tools

HTTP API

GET  /health                          POST /api/agents/{id}/execute
GET  /api/agents                       GET  /api/agents/{id}/status
POST /api/agents/{id}/restart          GET  /api/tokens/report
GET  /api/workflows                    POST /api/workflows/{id}/execute
GET  /api/mcp/servers                  GET  /api/mcp/tools
GET  /ws   (WebSocket streaming)

Examples

Runnable, mock-LLM (no keys needed) — see examples/: research-assistant, code-reviewer, customer-support.

Build from source

git clone https://github.com/axocoatl/axocoatl
cd axocoatl
cargo build --release          # binary: target/release/axocoatl
cargo test --workspace         # 340+ tests

License

Apache-2.0 — see LICENSE. Changes: CHANGELOG.md.

Ver o original no GitHub

Bom para

Relacionados