Guide Personal Agents · Chapter 8

    Running OpenClaw and Hermes on your own hardware.

    A gateway takes messages from Teams or Signal, tools run in a sandbox, a local model reads private notes, and an audit checks every change you make.

    As of: October 2026 · Editor: Marcel Haas

    • OpenClaw gateway
    • Ollama
    • Docker
    • Sandbox
    • Pairing
    • Security audit
    01

    What you need before the first command runs.

    At the start there is a list of five lines that Reto Kunz fills in for Muster AG within an hour, before the first installation. A machine that runs continuously: a small computer in the office, a server, a virtual machine at your hosting provider or a container on a server that already exists, plus a person who owns it. Model access: a strong cloud model for mail, planning and everything with tools, through an API account of the company with a project of its own, a spending cap at the vendor and keys kept apart from the configuration, and a local model through Ollama for private notes and drafts. The channels: Microsoft Teams first, because sign-in and policies are already in place there and IT approves the app in the tenant, then Slack, Telegram, WhatsApp or Signal; every channel is an entry point for messages from outside, so only the agreed ones. The identity: an account of its own for the agent with delegated permissions in mail, calendar and storage, as chapter 7 describes. And the decision who shares a gateway. Hermes Agent runs on the same list, locally or in a container, with any model, including a local one.

    02

    The gateway is the trust boundary.

    On the machine a gateway process runs that serves all channels, routes every message to the right session and drives heartbeat and schedules; the configuration sits in one file, the Control UI shows chats, sessions and settings in the browser, and set up as a service it starts with the machine. OpenClaw assumes one trust boundary per gateway: one person or a team that trusts each other; a shared gateway is not meant for people with opposing interests. The cut follows from that: one gateway for Anna Meier and her agent, a second for the sales team, never one for the whole company. Company and private agents we separate through distinct agents with their own workspaces, and every session has its own transcript. On a regular host the gateway binds to the local address; container images expose it by default, so there we always set token authentication, and we protect the configuration folder like credentials. Unknown senders receive a pairing code, and their message waits until you confirm them; in the trial we switch to a fixed list per channel, in groups only people on the group list can trigger the agent, and the codes expire after one hour.

    03

    SOUL.md, USER.md and MEMORY.md live in your workspace.

    An agent's workspace is the handover in file form. SOUL.md says how it works and where it stops, USER.md who you are and how you work, MEMORY.md what it has learned along the way, as short dated entries and a log for each day; skills sit next to them as folders with a SKILL.md and are loaded when they come up. All of it is text that the person reads, changes and deletes, and that is the strength of self-hosting: what the agent knows sits in files on your hardware and not in an account at the vendor. The folder belongs to the backup, and a copy in the person's own storage, OneDrive for example, moves along when the agent is a different one some day. Hermes Agent works on the same pattern, derives skills from completed tasks and remembers across sessions; what its files are called and what they hold we check with the vendor before we record it on the sheet per agent. What never belongs in the workspace: API keys, passwords, tokens. Those sit apart, with a spending cap at the vendor, and belong to the company.

    04

    Tools run in a sandbox, and tool rules block the rest.

    The gateway runs on the machine, the agent's tools in a Docker or Podman container, one sandbox per agent or per session. A single setting decides whether it sees the workspace not at all, read-only or read-write; additional folders OpenClaw checks against a denylist, system folders and the Docker socket stay out, and in the trial access is read-only. A deny list takes tools such as schedules or spawning further sessions away from the agent, commands outside the sandbox run only through elevated execution for few accounts, each with a prompt, and if the prompt reaches no one, the answer is no. Commands on paired devices are blocked by a gateway rule. Skills and plugins from ClawHub we treat like third-party code: an install policy decides before every install whether it is allowed, allowed with a warning or blocked, and our rule admits only entries from a reviewed list. The documentation recommends for agents with tools the strongest current model, a sandbox for sensitive tools and web search switched off where it is not needed, because smaller models are more susceptible to tool misuse and hijacked instructions. That is how an agent reads outside mail without a hijacked instruction ever reaching a command on your server.

    05

    The cloud model for tools, the local one for private matters.

    OpenClaw talks to cloud models through their APIs and to local models through Ollama or other model services; if the first model fails, the next one in the list steps in. For mail, planning and everything with tools the agent works with the strongest current cloud model and a second one as fallback; private notes and first drafts are taken over by an open model running locally through Ollama, so that text never leaves the machine. The gateway connects Ollama through its native interface, because tool calls break over the compatible address with /v1 and the model emits them as text. A local model needs enough graphics memory and context for agent work; we measure the hardware on your test set before it is bought, and the model server runs on the same machine or on one with a graphics card in the same network, reachable only for the gateway. Which model may see which data is on the sheet per agent, and what happens to the data the cloud model sees the OpenAI guide records: the vendor has not used API data for training since 1 March 2023, abuse monitoring logs are kept for up to 30 days, and the region is set per project.

    06

    You answer approvals in the Control UI or in the apps.

    At 17:40 the agent wants to install a package on the server to produce a PDF; the command appears in the Control UI, the operations owner reads it, approves it once, and the log records both. Sending, accepting, buying and deleting the agent prepares and then waits, and the request reaches you in the Control UI in the browser, in the macOS app or on your phone through the iOS and Android nodes, which you pair yourself and can remove at any time. Through the apps the agent can also take photos, read the screen or speak; commands on the devices stay blocked. Whoever is on holiday names a deputy in the directive, otherwise the draft stays a draft. The requests are the tool with which you train the agent: what you answer with yes three times the pilot can turn into a permitted action, what you refuse twice belongs in SOUL.md as a rule.

    07

    Audit, updates, backups, monitoring, and who is on call.

    After every change openclaw security audit runs and shows whether the configuration has drifted from the agreed rules; the result goes to the log, findings are fixed or justified. Updates arrive in a window the operations owner knows: read the release notes, take a backup, update the gateway, run the audit again, run a test with test data. Telemetry is limited to a version check that can be switched off, and there is no paid tier, so the upkeep rests with you or with us. Workspace and configuration are backed up, the configuration encrypted like credentials, with a restore that has been rehearsed once. Monitored are the gateway process, the space on the disk, the reachability of the models, open pairing requests, refused approvals and the rhythm that rests outside working hours. And one person is on call, in your IT or in our support and operations team, with a procedure for the incident that starts with three moves: disable the identity, stop the gateway, replace the keys.

    08

    On a rented server, stricter rules apply.

    A virtual machine at a hosting provider sits on the internet, and that changes the order of the moves. The gateway binds to the local address only, you reach the Control UI through an SSH tunnel or behind a reverse proxy with sign-in, and a token protects it in every case. The firewall admits only the connections the channels need, the gateway runs under a user of its own without administrator rights, the Docker socket stays out of the agent's reach, the disk is encrypted, keys sit in a secrets store, and sign-in to the server works with a key pair only. The model server listens on its own machine only or on the private network between gateway and model machine. The audit runs before the first channel, not after the first incident, and the session transcripts are copied daily into the company storage, so that a compromised server does not take them along. For the computer in the office the same applies on a smaller scale: local address, a user without admin rights, updates in the window, and nobody closes the lid in the evening. The detailed procedure is in the OpenClaw guide below, into which the earlier guide for the rented server was merged.

    What you decide.

    A computer in the office, a virtual machine at the hosting provider or a container on an existing server?

    Options

    1. 1A small computer in the office
    2. 2A virtual machine at your hosting provider
    3. 3A container on a server that already runs

    Our recommendation: For the trial a virtual machine with Docker on which nothing else runs, with a token for the Control UI and the audit before the first channel; for operation afterwards the place where your IT maintains servers anyway. One machine per gateway, that is per trust boundary, and the size according to the model that computes locally. A computer in the office fits when it runs continuously and someone owns it; a rented server fits when the rules of the last section apply.

    What we advise against: The person's laptop as the gateway, closed in the evening, or a server that nobody owns.

    Who shares a gateway?

    Options

    1. 1One person per gateway
    2. 2A team that trusts each other
    3. 3The whole company on one gateway

    Our recommendation: One person or a team that trusts each other, as the documentation intends, with distinct agents and distinct workspaces for company and private matters. A sales team with shared quotes can share a gateway, management and apprentices cannot, and people with opposing interests never. More gateways cost more machines and operations; fewer gateways cost the separation, and that cannot be retrofitted.

    What we advise against: One gateway for everyone, with a shared memory in which management's mail sits next to the apprentices' mail.

    Does the sandbox see the workspace read-only or read-write?

    Options

    1. 1Not at all
    2. 2Read-only
    3. 3Read and write

    Our recommendation: Read-only in the trial, and afterwards read-write only for the agent's own folder once the pilot has shown what it needs it for, such as notes and drafts it files itself. System folders, the Docker socket and the folders on the denylist stay out in every case. What the agent gives to the outside does not go through the sandbox but through the approval.

    What we advise against: Tools without a sandbox directly on the server because Docker happened not to be installed.

    Who runs it after the pilot?

    Options

    1. 1The person herself
    2. 2A named person in your IT
    3. 3Our support and operations team

    Our recommendation: A named person in your IT or our support and operations team, with a clear split: the person maintains SOUL.md, skills and memory and answers requests; operations keeps updates, audit, backups, monitoring and the on-call duty. The hours for that are in the pilot's operations column, and they belong to the bill from chapter 6.

    What we advise against: Nobody. An agent that nobody updates keeps running with an old gateway and old rules until someone notices.

    Status and sources.

    As of: October 2026

    • As of October 2026

      Open source, backed by a foundation.

      OpenClaw is MIT-licensed and developed by the OpenClaw Foundation, an independent US non-profit. There is no paid tier, and telemetry is limited to a version check that can be switched off.

      Source: OpenClaw Docs
    • As of October 2026

      A gateway is a trust boundary.

      OpenClaw assumes one trust boundary per gateway: one person or a team that trusts each other. A shared gateway is not meant for people with opposing interests. On a regular host the gateway binds to the local address, and unknown senders receive a pairing code.

      Source: OpenClaw Docs: Security
    • As of October 2026

      Prompt injection is not solved.

      The documentation states that smaller models are more susceptible to tool misuse and hijacked instructions. For agents with tools it recommends the strongest current model, a sandbox for sensitive tools and web search switched off where it is not needed.

      Source: OpenClaw Docs: Prompt injection
    • As of October 2026

      Connect Ollama through its own interface.

      OpenClaw talks to Ollama through its native /api/chat interface. Through the OpenAI-compatible address with /v1, tool calls break and the model emits them as text.

      Source: OpenClaw Docs: Ollama
    • As of October 2026

      API data: no training, a Europe region per project.

      OpenAI has not used API data for training since 1 March 2023, and abuse monitoring logs are kept for up to 30 days. The Europe region, which OpenAI labels EEA and Switzerland, stores and processes in the region for the listed endpoints and requires a retention option approved by OpenAI. The Agents API is not listed there.

      Source: OpenAI: Data controls in the OpenAI platform
    • Since April 2026

      Entra Agent ID, one identity for every agent.

      Microsoft Entra Agent ID is generally available: an identity and authorisation framework for AI agents based on OAuth 2.0, MCP and A2A, for agents from Microsoft Foundry, Copilot Studio and third-party platforms. Access, conditions and logs therefore apply per agent as they do for people.

      Source: Microsoft Learn: What is Microsoft Entra Agent ID

    Frequently asked questions.

    Which machine is enough to start with?

    For the gateway alone a small computer that runs continuously and runs Docker or Podman; the size is set by the local model, because agent work needs enough graphics memory and context. For the trial we take a virtual machine at your hosting provider on which nothing else runs, and we measure the local model on your test set before hardware is bought. The person's laptop is no gateway: it is closed in the evening, and the rhythm with it.

    Does the gateway have to be reachable from the internet?

    The Control UI does not. It stays on the local address and is reachable through an SSH tunnel or behind a reverse proxy with sign-in, with a token in every case. Which connections the individual channels need is in the documentation of each channel, and the firewall admits exactly those and nothing else. Container images expose the gateway by default; that is why token authentication is the first move there and the audit the second.

    What happens when the server goes down?

    Then the agent stands still: the rhythm rests, messages in the channels stay unread, open requests stay open, and because an unanswered request counts as a no, nothing happens. Set up as a service the gateway restarts with the machine; workspace and configuration come back from the backup, whose restore has been rehearsed once. We test the restart after an outage in the trial on purpose, with test data, before real accounts are connected.

    How does an update reach the gateway without breaking anything?

    In a window the operations owner knows, and in a fixed order: read the release notes, back up workspace and configuration, update the gateway, run openclaw security audit, run a test with test data, and on a finding go back to the backup. OpenClaw's version check reports new versions and can be switched off; when to update is your call. Skills and plugins we update like third-party code, read before they go live.

    Can the agent run commands on the server?

    In the sandbox yes, on the server itself only through elevated execution, which few accounts have and which triggers a prompt every time; if the prompt reaches no one, the answer is no. Tools such as schedules or spawning further sessions the deny list takes away from the agent, and commands on paired devices are blocked by the gateway rule. An agent that reads outside mail therefore reaches no command on your server, even when a mail asks it to.

    Does Hermes Agent run under the same rules as OpenClaw?

    Under the same house rules, because the rules belong to us and not to the agent: an identity of its own, approvals before sending, booking, paying and deleting, sandbox, log, audit. Hermes Agent brings mechanisms of its own, skills from completed tasks, a memory across sessions, any model, including local ones; how it sets them and where it stores them we check with the vendor before the start. For the trial it gets a bounded toolset, stop conditions and an evaluation of what it has learned.

    What does the model vendor see when the agent calls a cloud model?

    What the agent sends it: its instructions, excerpts from the memory and the mail and documents that belong to the task. That is why the sheet per agent says which data the cloud model may see, and private notes go to the local model. OpenAI has not used API data for training since 1 March 2023, abuse monitoring logs are kept for up to 30 days, and the region is set per project; the details are in the data boundaries chapter of the OpenAI guide, and for other vendors we check the terms before connecting them.

    Next chapter · Chapter 9

    Who does what so that your agents stay in operation.

    Who does what, so that personal agents stay in operation?