In 2026 the open-source agent ecosystem split into three layers: gateway (OpenClaw), self-improving runtime (Hermes), multi-agent orchestrator (Paperclip). Onepilot deploys all three to a host you can SSH into and makes the iPhone the supervisor.
The three layers compose on one host. OpenClaw routes channels and plugins. Hermes executes and learns skills. Paperclip coordinates multiple agents into a virtual company. NousResearch ships a hermes-paperclip-adapter so a Hermes runtime can be hired as a managed employee inside a Paperclip company. iOS does not allow long-running background processes, so all three live server-side; the iPhone is the supervision surface.
| Layer | Agent | Role | Not its job |
|---|---|---|---|
| Gateway | OpenClaw | Channel routing, plugin dispatch, credentials | Not the brain. Not the org chart. |
| Runtime | Hermes | Self-improving skill execution, persistent memory | Not a chat client. Not a coordinator. |
| Orchestrator | Paperclip | Multi-agent org chart, ticketing, budget governance | Not an agent. Not a runtime. |
OpenClaw (gateway) — A long-running process on your host that routes between LLM providers, channel plugins (Telegram, Discord, Slack), and skill files. The gateway is the agent's front door.
Hermes (runtime) — An always-on agent process that auto-generates reusable skills from experience and recalls past work via persistent memory. The runtime is the agent's brain.
Paperclip (orchestrator) — A control plane that hires multiple agents (Claude Code, Codex, OpenClaw, Hermes, anything on OpenRouter) into roles inside a virtual company, with goals, heartbeats, and budget caps. The orchestrator is the company.
The three layers run side by side on one host. A Telegram message arrives: OpenClaw's channel plugin receives it and dispatches to the registered agent. A Hermes runtime picks up the task, recalls relevant past skills from its FTS5 memory, executes, and writes a new skill file if it learned something. If the host is running Paperclip, that Hermes runtime is hired into a role inside a company, and the work is tracked as a ticket against a budget. The phone shows what happened — message log, skill diff, ticket status — but never holds the running process.
Run any one alone. The minimum useful stack is just OpenClaw on a Raspberry Pi. Add Hermes when you want skill memory. Add Paperclip when you have more than one agent and need a chain of command. The stack grows by adding adapters, not rewriting layers.
You want one agent reachable from Telegram, Discord, and Slack with the same memory.
Start with OpenClaw →You want the agent to learn reusable skills from your work and not lose them between sessions.
Start with Hermes →You have multiple agents already and want roles, tickets, and a budget keeping them in line.
Start with Paperclip →
Personal AI assistant with a Gateway-and-nodes architecture.
Onepilot provisions the OpenClaw Gateway on any host you SSH into — your Mac mini, a Linux box at home, a Windows machine, a Raspberry Pi, or a $5/month VPS — and pairs the iPhone as a node. The Gateway lives wherever you put it; the iPhone is the supervisor.
Self-improving agent from NousResearch — Linux/macOS/WSL2/Termux only, until now.
Hermes does not officially support iOS. Onepilot deploys the Hermes runtime to any Linux or macOS host you can SSH into — a Mac mini at home, a Linux box, a Raspberry Pi, or a remote VPS — and drives it from iPhone via Telegram, Discord, or Slack.
Open-source orchestrator that turns Claude Code, Codex, and other agents into a coordinated team.
Paperclip is not an agent — it's the coordinator. A CEO process delegates goals to engineer and QA agents (Claude Code, Codex, OpenCode, anything on OpenRouter), with org charts, ticketing, and governance baked in. Onepilot deploys the Paperclip control plane on any host you SSH into and gives you the iPhone supervisor on top — so you can spin up a multi-agent company from your phone and watch the work move across roles.
Anthropic's coding agent — already available through the Onepilot terminal.
Claude Code is a terminal CLI. Because Onepilot already includes a real PTY terminal, Claude Code works out of the box: SSH into your host, run claude, use it interactively. No wizard step needed. The value Onepilot adds is the layer around the shell — file browser, git, cron, channel routing, and an iOS Keychain-backed API key.
OpenAI's coding agent — already available through the Onepilot terminal.
Codex CLI is a terminal CLI like Claude Code. It works today inside the Onepilot SSH session — install it once on your host, then run codex from any PTY tab. iOS Keychain holds your API key; the file browser, git tab, and cron tab wrap around the shell.
A personal agent stack is a self-hosted set of three layers — gateway, runtime, and orchestrator — that together let one operator run AI agents persistently. In 2026 the canonical open-source choices are OpenClaw (gateway), Hermes Agent (self-improving runtime, NousResearch), and Paperclip (multi-agent orchestrator). The three install on the same host and compose: the gateway routes input, the runtime executes, the orchestrator coordinates multiple runtimes.
They sit at three different layers of the same stack. OpenClaw is the gateway — it routes channels, plugins, and skills. Hermes is the runtime — a self-improving agent process with memory. Paperclip is the orchestrator — it coordinates multiple agents into a virtual company. You can run any one alone, or all three together on the same host. NousResearch ships a hermes-paperclip-adapter that lets a Hermes runtime be hired as a managed employee inside a Paperclip company.
Pick OpenClaw if your bottleneck is channels (Telegram, Discord, Slack) and plugin routing. Pick Hermes if your bottleneck is the agent itself learning skills from your work. They are not direct substitutes — OpenClaw is a gateway, Hermes is a runtime. A common 2026 setup runs both on one host: OpenClaw handles the chat surface, Hermes handles execution and memory. Read the comparison: /blog/hermes-vs-openclaw.
Hermes officially supports Linux, macOS, WSL2, and Termux on Android. iOS is not on that list. Onepilot deploys Hermes to a Linux or macOS host you can SSH into — a Mac mini, a Raspberry Pi, a NAS, or a small VPS — and drives it from iPhone via Telegram, Discord, or Slack. The agent runs on the host; the iPhone is the supervisor.
Yes, through Onepilot. Paperclip's control plane (Node.js + React dashboard) runs on a host you SSH into. Onepilot's iPhone app provisions the host, hires the first CEO agent, and exposes the dashboard, heartbeat status, and budget approvals through a mobile surface. The orchestrator runs server-side; the iPhone is where you approve, pause, and audit.
The iPhone is the supervision surface, not the runtime. iOS sandboxes prevent persistent background processes, so the gateway, runtime, and orchestrator all live on a host you own or rent. The phone is where you watch the agent work, approve actions, edit skill files, and route messages — a control room, not a server. This is the gap Onepilot fills: the phone-side IDE for an agent stack hosted somewhere else.
OpenClaw is not a brain — it does not learn skills or coordinate other agents. Hermes is not a chat client — it does not handle Telegram or Discord routing on its own. Paperclip is not an agent — it has no inference loop of its own; it hires real agents (Claude Code, Codex, Hermes, OpenClaw, OpenRouter models) and assigns them work. Mixing the layers — running Paperclip without an agent to hire, or Hermes without a channel — is the most common 2026 setup mistake.
One Linux host with 2 GB RAM and 20 GB disk handles the gateway plus a single Hermes runtime. Add 2 GB RAM per additional concurrent agent inside Paperclip. Common 2026 setups: a Raspberry Pi 5 (8 GB) at home for hobby use, or a small VPS for always-on work. Onepilot is free; you bring the host and the LLM keys.
Onepilot is free to get started. Bring your own host, your own LLM key, and the iPhone.
Join the waitlist