arizuko › comparable systems
comparable systems
The systems that overlap with arizuko on channels, tenancy, isolation or memory, and how each one differs.
Each comparable system does a part of the job
Each line says how the system differs from arizuko. The agents hub on krons at https://krons.fiu.wtf/pub/krons/agents/ reads the source of many of them, and https://krons.fiu.wtf/pub/krons/agents/comparison/ compares five side by side. krons at https://krons.fiu.wtf is the reference deployment of arizuko.
Personal agent gateways serve one operator
| System | Address | How it differs |
|---|---|---|
| OpenClaw | https://github.com/openclaw/openclaw | A personal assistant on many chat platforms from one Node.js process. An agent is a row in its config. A Docker sandbox for tools is optional, and no VM runs a turn. |
| NanoClaw | https://github.com/nanocoai/nanoclaw | A small rewrite of the OpenClaw model. The Claude Agent SDK runs in one container per session, and a per-session SQLite file is the only channel to the host. arizuko runs each turn in its own KVM guest and reaches the host through an MCP socket. |
| Muaddib | https://github.com/pasky/muaddib | A chat assistant for teams that runs tool calls in QEMU micro-VMs. The model SDK and the web tools run on the host, and all users of a channel share one VM. |
| ElizaOS | https://github.com/elizaOS/eliza | A framework with character.json personas and a plugin marketplace. Agents run in the host process with no sandbox, and memory is PostgreSQL with pgvector. |
| Milady | https://github.com/milady-ai/milady | ElizaOS as a desktop and mobile app for one person, local first. |
| Hermes Agent | https://github.com/NousResearch/hermes-agent | A long-running Python agent that writes its own skills and scans skills for threats. One process keeps a session across many turns, with no VM per turn. |
| takopi | https://github.com/banteg/takopi | A Telegram bridge to the coding CLIs Codex, Claude Code, OpenCode and Pi, with one git worktree per branch. One operator and no tenants. |
| Letta | https://github.com/letta-ai/letta https://github.com/letta-ai/letta-code | A server for agents that keep memory across runs, local or in Letta Cloud. letta-code sends replies only through a message tool, for one operator with no tenants. An arizuko turn delivers its final text block. |
| brainpro | https://github.com/jgarzik/brainpro | A Rust coding assistant that sends each kind of task to a different model provider. A chat pairs to a session with a code, and no sandbox runs the tools. |
| HOME23 | https://github.com/notforyou23/home23 | An agent that runs all the time in continuous loops. An arizuko agent runs only during a turn. |
Team harnesses keep a sandbox per scope or per thread
| System | Address | How it differs |
|---|---|---|
| qm | https://github.com/yc-software/qm | A multiplayer agent harness in Slack and on the web, hosted in the cloud account of the team. It keeps one durable sandbox per scope, filters egress with Envoy and uses MCP servers without serving one. Its scopes are five flat kinds, where the path of an arizuko folder has no depth limit. |
| Centaur | https://github.com/paradigmxyz/centaur | Slack agents in one Kubernetes pod per thread, a workflow engine that replays steps, and a proxy that adds credentials on the wire. arizuko has no workflow engine. routd runs a turn that fails without a reply again, up to MAX_TURN_RETRY times, default 3. |
| OpenClaw Managed Agents | https://github.com/stainlu/openclaw-managed-agents | An open implementation of the Anthropic Managed Agents API over one Docker container per session. The API has four nouns: agent, environment, session and event. |
| Cloudflare Agents | https://github.com/cloudflare/agents | An SDK for agents that run as Durable Objects on Cloudflare, each with its own SQLite state. Cloudflare runs the platform, where arizuko runs on a host you own. |
| Dust | https://dust.tt/ | A hosted service for agents over company data, with many data connectors. arizuko has no connector sync and is not a hosted service. |
| Zapier Agents | https://zapier.com/agents | Hosted agents that act through the app integrations of Zapier. |
Frameworks and builders leave the hosting to you
| System | Address | How it differs |
|---|---|---|
| LangGraph | https://github.com/langchain-ai/langgraph | A library for agent programs as graphs. You write the service, the channels and the tenancy. |
| AutoGen | https://github.com/microsoft/autogen | A framework from Microsoft for conversations between several agents in one program. |
| CrewAI | https://github.com/crewAIInc/crewAI | A Python framework for crews of agents with roles that work on one task, and a hosted enterprise product. |
| smolagents | https://github.com/huggingface/smolagents | A small Python library in which the agent writes its actions as code. |
| Flue | https://github.com/withastro/flue | A TypeScript agent framework that runs tools in an in-memory virtual shell by default. A container is an option, not the default. |
| Dify | https://github.com/langgenius/dify | A visual builder for LLM apps, workflows and retrieval. An agent is a node in a graph, not a folder. |
| Flowise | https://github.com/FlowiseAI/Flowise | A visual builder for LLM flows. |
| n8n | https://github.com/n8n-io/n8n | Workflow automation with AI agent nodes. An agent is a step in a workflow. |
Adjacent tools share one part of the shape
| System | Address | How it differs |
|---|---|---|
| Aeon | https://github.com/aeonfun/aeon | Agents that run skills on GitHub Actions on a schedule and commit their state to git. A run is a scheduled job, not an answer to a chat message. |
| Autolith | https://github.com/lambda-symbolics/autolith | A Common Lisp agent that changes its own running image and can roll it back. arizuko keeps no live image: each turn boots a new guest. |
| Entire | https://entire.io/ | A git host that links each agent session to the commit it made. arizuko is not a git host. |
| Onyx | https://github.com/onyx-dot-app/onyx | Self-hosted search and chat over company documents, with connectors. It does the retrieval that arizuko does not do. |
Four limits of arizuko are open
| Limit | What is true now |
|---|---|
| The runed container is not locked down | Compose runs runed as root with NET_ADMIN, SYS_ADMIN and apparmor=unconfined, because the default AppArmor profile denies the mount(2) calls of the staging tree. The turn itself runs in a KVM guest. No narrower AppArmor profile for runed is written. |
| No cluster has run the Helm chart | The chart in deploy/helm/arizuko/ passes helm lint and helm template. It runs routd and runed as one replica each on the node that holds the data volume. |
| No trace viewer | Every daemon exports OTLP logs and spans when OTEL_EXPORTER_OTLP_ENDPOINT is set. The collector and the viewer belong to the operator. |
| No loop judges finished turns | anteval runs eval suites when an operator starts it. No daemon scores a turn after the turn ends. |
Design directions for arizuko live in the specs/ directory of the repository. A spec that is folded into the docs is deleted and stays in git history: git log --diff-filter=D --name-only -- specs/.
Sources for arizuko: [README.md] [routd/loop.go] [groupfolder/folder.go] [ant/CLAUDE.md] [compose/compose.go] [deploy/helm/arizuko/README.md] [obs/README.md]. The other systems are described from their repositories and the krons hub. This page does not re-measure them.