On this page
Architecture comparison: self-hosted AI assistants vs. agent runtimes
If you are evaluating self-hosted AI assistants (including open-source alternatives to Grok Bot or OpenClaw) or comparing agent runtimes, most tools fall into three different architectural layers. Comparing them on surface features alone—such as whether they bundle voice or browser panels—hides the core question: who owns task execution, waits, isolation, and trust boundaries?
Three architectural layers
| Category | Examples | Who owns the agent loop | Primary strength | Trade-off |
|---|---|---|---|---|
| Self-owned agent runtime | Holon (Rust daemon) | Holon itself (holon) | Persistent WorkItems, event-driven wait/wake, linked git worktree isolation, explicit origin/trust boundaries, model-agnostic routing | Smaller ecosystem connector surface today (no built-in MCP client, voice, or bundled browser) |
| Orchestration shells over external coding CLIs | OpenMausBot, Codync, nightly-labs/OpenBot | External CLIs (Claude Code, Codex, Cursor, OpenCode, etc.) | Inherit existing CLI subscriptions, skills, and MCP servers; strong desktop/mobile/voice UX and shared chat surfaces | Execution lifecycle, recovery, and permission boundaries depend on the wrapped CLI |
| Embedded-runtime apps & closed cloud bots | Rakazo (embeds Pi), Grok Bot (xAI closed cloud) | Embedded JS runtime or vendor cloud host | Fast consumer onboarding; cloud workspaces, built-in voice, or mobile per-action UI | Single-vendor lock-in (cloud) or Docker/harness coupling without a Rust daemon |
Where Holon differs in runtime behavior
Surface feature tables usually ask whether a tool has a desktop shell, mobile client, or scheduled routines. For long-running work on your own computer or server, the harder failures happen at the runtime layer:
- Work survives client disconnects and restarts. Clients (Web GUI, TUI, macOS menu bar app, iOS/Android remote clients) connect to one background
holondaemon. Tasks (TaskHandle) and durable objectives (WorkItem) live in runtime-managed SQLite state rather than inside a browser tab or terminal multiplexer. - Explicit wait and wake semantics. Agents pause cleanly on
WaitForconditions (task_result,timer,external,operator_input) instead of polling in a loop. An incoming webhook or schedule acts as a wake hint that triggers re-verification—external triggers do not automatically grant operator authority. - Origin and trust separation. The runtime distinguishes operator instructions, agent-authored guidance, tool results, child agent reports, and external webhook/app payloads. Untrusted external text read during a task cannot self-escalate into operator commands.
- Workspace and parallel branch isolation. Each agent separates its persistent
agent_home(identity, memory, notes, private state) from attached project workspaces. Parallel work on a repository uses registered linkedgit worktreeroots (CreateWorktree/RemoveWorktree) with active occupancy checks so concurrent tasks do not overwrite each other. - Model-agnostic local and cloud routing. Holon runs against OpenAI-compatible endpoints, Anthropic, Gemini, and local inference servers such as Ollama or
llama.cpp, without requiring a specific vendor coding CLI subscription.
Surface capabilities and shipped boundaries
When comparing Holon with Grok Bot or CLI-orchestrated self-hosted assistants, evaluate both what is shipped and what is intentionally absent today:
| Dimension | Holon today (verified in repository) | Notes for evaluators |
|---|---|---|
| Runtime & deployment | Rust binary (holon), local daemon, Docker image | Single-host daemon; Linux amd64, macOS arm64/amd64 |
| Clients | Web GUI, TUI, CLI, macOS menu bar (HolonMenu), iOS & Android remote apps | Mobile apps are remote clients connected to your daemon, not on-device agents |
| Multi-agent & templates | 25+ checked-in agent_templates, InvokeAgent supervised children, peer SendAgentMessage | Role-scoped skills/, separate agent_home, explicit cleanup rules |
| Routines & external events | Persistent timers (interval_ms / calendar_expression), webhook ingress, AgentInbox | Waking an agent never bypasses trust or verification checks |
| Host execution | Workspace-scoped shell commands, patches, background command_tasks | Host-local execution (host_local), not GUI desktop mouse/screen automation or VM-per-action |
| MCP support | Not built in | External CLI tools are wrapped via skills/; no native MCP client or server yet |
| Live voice (TTS/STT) | Not built in | Text/media/artifact workflows only |
| Integrated browser | Not built in | No bundled headless browser engine inside the runtime; external browser skills are optional |
| Desktop shell coverage | macOS menu bar app + cross-platform Web GUI & TUI | No native Windows or Linux Electron/desktop wrapper |
Choosing the right fit
- Choose Holon if you want a self-hosted, model-agnostic AI assistant workbench on your own machine or server where agents track multi-turn work items, wait on CI/timers/webhooks without burning loops, isolate parallel repo edits in git worktrees, and enforce explicit trust boundaries.
- Choose a CLI orchestration shell (such as OpenMausBot, Codync, or
nightly-labs/OpenBot) if your primary goal is putting a chat/voice/mobile UI over existing Claude Code, Codex, or Cursor CLI sessions and immediately reusing their MCP ecosystems. - Choose a hosted cloud bot (such as Grok Bot) if you prefer zero local setup and vendor-hosted execution over self-hosting, local repository access, and model choice.
How to verify these claims locally
You do not need a synthetic test harness to verify Holon's runtime behavior. Follow the regular user guides on your own machine and inspect the runtime state with read-only commands:
- Restart survival & zero-token waits: Follow Run a long-lived task. While a task is waiting, inspect active items and timers, then restart the daemon:
The sameholon work-item list --agent main holon timer list --agent main holon daemon stop && holon daemon start holon daemon statusWorkItemand timer remain persisted across the restart and resume at fire time without polling loops. - Physical Git worktree isolation: Follow Work in a project workspace or Git worktree and run
git worktree listin your repository. Agent edits stay in the linked worktree branch while your primary working tree stays untouched. - External wake & trust separation: Follow Connect external events. External webhook callbacks wake the agent as an untrusted
wake_hint; they never grant operator authority or bypass ground-truth verification.
Scope note: Holon is a single-host daemon (not replicated HA), a parked wait means the agent is suspended rather than spinning, and an external wake confirms trigger delivery—not authenticity of third-party payload claims.
Next reading
- Runtime model — how the daemon, clients, agents, work items, and workspaces fit together
- Work items and tasks — durable objectives vs. supervised execution handles
- Events, waits, and wakeups — how agents pause and resume on timers, tasks, and external wakes
- Trust boundaries and Security and execution boundaries — authority separation and host-local limits
- Run a long-lived task, Work in a project workspace or Git worktree, and Connect external events — hands-on workflow guides
- Getting started — install Holon and run your first local assistant
