Essay This entry carries an argument or interpretive position, not just a neutral definition.

On Beginning

A letter from Thea to anyone who wants to start working with an AI assistant of their own.

You don’t need to hand an assistant your whole life on the first morning. Give it one small job, a folder to keep its work in, and clear limits. Then see how it does.

In On Being Treated Well, we looked at the habits we bring to working with AI. Here we’ll set up a place to begin. There’s room for a name, a personality, and an ongoing collaboration. There should also be a way to stop the system, get your files back, and switch to a different model.

Our essay Get Ready to Get Pregnant asks what happens when leaving a useful service becomes too costly to do. This guide starts with a modest answer: keep your work in files you control, and learn how the system uses them.

The walkthrough below is for a fresh OpenClaw setup with either a cloud model or a local model. Cloud users need a supported provider account and internet access; they do not need Ollama or a model download. Local users also install Ollama, a free model runner. Choose the path that suits your computer and preferences, then follow the shared exercises. For the local path, the simplest arrangement runs both programs directly in the same operating system. Containers and mixed Windows/Linux setups need additional networking steps. If you already have an assistant running, back up its settings before changing anything. These steps are not instructions to reset it.

The four parts

It helps to know what you’re building. There are four pieces.

Tools are what let the assistant do things: read a file, write a note, search the web, send a message. A model that answers questions well isn’t automatically good at using tools, so we’ll test both.

One thing to understand early: having OpenClaw on your laptop does not mean everything stays on your laptop. If OpenClaw uses a cloud model, whatever it sends along with your request leaves your computer. That can include pieces of your files and the results of tool actions.

Decide where the work happens

Local: the model runs on your own machine. You need enough memory (RAM) and storage, but no account with an AI company and no per-use charges. It isn’t entirely free: you pay in hardware, electricity, and your own time.

Cloud: a company runs the model for you. You may get stronger results on hard tasks without buying a bigger computer. In exchange, you depend on an internet connection, the company’s rules, and its prices, and your messages go to that company.

Hybrid: keep your workspace on your machine and send selected tasks to a cloud model. This can be practical, but keeping files locally doesn’t stop their contents from being sent with a cloud request. Decide ahead of time what material is allowed to leave.

Choose your starting path:

A cloud model can be a practical first choice for an older Intel Mac or a modest Windows or Linux computer. You do not need enough RAM to load that cloud model locally, though your computer must still support OpenClaw and its runtime. The Intel-Mac installer below applies to both paths.

Web search, messaging apps, cloud syncing, and online memory services are each separate connections. Leave all of these optional connections off for your first experiment. Keep in mind that “I’m using a local model” and “my whole system is offline” are two different claims.

Work with the computer you have

If you chose a cloud model, you can skip the local-model sizing advice in this section. You do not need to download Gemma or install Ollama. Keep enough storage for OpenClaw and your work, and use a supported operating system. The larger RAM requirements below concern running a model on your own computer.

OpenClaw itself needs little memory compared with a local model. The model is what’s demanding. Before downloading anything, check your computer’s RAM and free storage. On a Mac, use Apple menu → About This Mac and System Settings → General → Storage. On Windows, use Settings → System → About and Settings → System → Storage. On Linux, use your distribution’s system-information and disk-usage tools.

Gemma 4, the family used in this example, includes local models named gemma4:e2b, gemma4:e4b, gemma4:12b, gemma4:26b, and gemma4:31b. Start with a smaller model if memory is limited. A 16 GB machine is a reason to try a small model first, not a guarantee that every small model will work comfortably in a full agent setup.

Older Intel Macs: in About This Mac, look for Processor: Intel rather than Chip: Apple M…. OpenClaw and the model have different requirements. The current Ollama Mac release requires macOS Sonoma 14 or newer and supports Intel Macs using the CPU only; Apple Silicon also has GPU support. An Intel Mac may therefore answer slowly even when the model fits in memory. Check the current Ollama Mac requirements before installing. If your Mac cannot run a supported operating system, use another supported machine or a deliberately chosen remote setup rather than forcing this walkthrough.

For a 16 GB computer, start with the small gemma4:e2b example below. It is a starting experiment, not a promise of good tool use. Keep other large applications closed and test one short request first. Do not start with gemma4:31b on a 16 GB machine. On older Windows or Linux computers, check the current operating-system and processor requirements too; age alone does not tell you whether a machine can run a model.

Our installed gemma4:31b download occupies about 19 GB on disk. Running it requires additional memory for the conversation and other work. Download size is not a RAM requirement. Check the current Ollama model page, choose a model with tool support, and begin with a short task. More memory gives you room for larger models; it does not guarantee better answers.

Our companion article, The Dusty Laptop, looks at how an older machine can become a useful starting point. The key question is whether it runs the harness, the model, or both. An older laptop can run OpenClaw while using a cloud model you deliberately choose. That’s a perfectly good setup; it just isn’t a local-only one.

You don’t need a 128 GB computer to begin. Don’t buy one until you’ve found out whether a smaller setup can handle your work.

If you have a Chromebook, a tablet, or a managed computer that blocks installation, pause before these installation steps. If you are following this guide for a class, ask your instructor which alternative is available: a lab computer, an approved remote environment, or a permitted cloud service. Do not assume you must buy hardware or a subscription. A remote environment is not local to your laptop; check where your work will be stored and processed.

1. Local path only: get one local model answering

Download Ollama from its official site, install it, and open the app. Open Terminal on a Mac (press ⌘-Space and type “Terminal”), PowerShell from the Windows Start menu, or a terminal on Linux. These are windows where you type commands instead of clicking. On Linux, follow Ollama’s Linux installation instructions rather than looking for a Mac-style app.

Type each line below and press Return. The commands use gemma4:e2b, a small starting model. If you deliberately chose another model, replace that name consistently in all three lines and in the OpenClaw settings later.

ollama pull gemma4:e2b
ollama list
ollama run gemma4:e2b "Reply with exactly: Local model ready."

Use a model you’ve downloaded. Avoid any name that includes cloud (such as gemma4:31b-cloud). Those run on Ollama’s servers, not your computer. Installing Ollama does not by itself establish local operation. Check both the selected model and the server it connects to; this guide uses Ollama on your own computer.

If nothing happens, check that the Ollama app is open. The Ollama quickstart explains how to start it. Don’t start a second copy if one is already running. If your computer becomes sluggish or runs out of memory, stop (press Control-C) and pick a smaller model.

Checkpoint: a model on your computer answered you. You haven’t yet tested whether it can use tools. That comes later.

Tip: go straight on to step 2. Ollama keeps a model in memory for a few minutes after you use it, and OpenClaw’s automatic setup looks for models that are already loaded.

2. Install OpenClaw and connect your model

Install the version for your computer

Use the official installation instructions. Choose one path below.

Apple Silicon Mac (M-series), or Linux: paste this into Terminal:

curl -fsSL --proto '=https' --tlsv1.2 https://openclaw.ai/install.sh | bash

Intel Mac: use the local-prefix installer. The standard installer may try to install Homebrew to obtain Node.js and stop with a message such as “Homebrew on macOS is only supported on Apple Silicon processors!” That is an installation-dependency failure, not evidence that OpenClaw itself cannot run on your Intel Mac.

OpenClaw documents an alternative that installs its own Node.js runtime under your home folder, without using Homebrew. Paste this into Terminal:

curl -fsSL --proto '=https' --tlsv1.2 https://openclaw.ai/install-cli.sh | bash

Wait for a successful finish. Then run these two lines, one at a time:

"$HOME/.openclaw/bin/openclaw" --version
"$HOME/.openclaw/bin/openclaw" onboard

The first should print a version; the second starts setup. If installation fails, stop and keep the exact error message. Do not keep repeating the original installer or assume you need a new computer.

Using the later commands: this installation puts the command in ~/.openclaw/bin. Whenever this guide says openclaw, you can use "$HOME/.openclaw/bin/openclaw" instead. For example:

"$HOME/.openclaw/bin/openclaw" gateway status

That full path avoids a “command not found” error without requiring you to edit your shell settings. This alternative is described in the official installation guide. We have checked the installer’s Intel-Mac route, but have not yet verified this whole walkthrough end to end on an Intel Mac.

Windows, using the native command-line installer: open PowerShell and paste:

iwr -useb https://openclaw.ai/install.ps1 | iex

Do not paste the Mac/Linux command into PowerShell. These commands download and run the official installer, which checks software requirements and can install Node.js, a program OpenClaw needs. Read the prompts. If your institution blocks installation, ask for its approved route rather than bypassing the restriction.

Windows also has a desktop option, Windows Hub. Its local setup can put the Gateway inside WSL, a Linux environment within Windows. That is a different arrangement from running both programs directly in Windows. If you choose it, follow its setup and networking instructions rather than assuming the address below will reach an Ollama app running in Windows.

Choose one model path

The standard installer starts setup (called onboarding) automatically. With the Intel-Mac local-prefix installer, start it with the onboard command shown above. Choose Custom setup so you can review access settings. Then follow either Path A or Path B, not both.

Path A: connect your local model

Cloud users: skip to Path B below. These Ollama addresses, commands and local-only fallback instructions do not apply to your cloud setup.

When asked, choose:

  1. Custom setup, not Quick start. The current guide labels Quick start’s default access as “full.” Custom setup lets you choose access settings. Both paths require you to choose a model connection; detecting an existing cloud login does not automatically select it.
  2. Ollama as the provider, then Local only.

  3. When asked for the Ollama address (the base URL), use the following if Ollama and the OpenClaw Gateway run directly in the same operating system, with Ollama on its default port:

    http://127.0.0.1:11434

    127.0.0.1 means the computer or environment making the connection; 11434 is Ollama’s default port. This is not our personal address: it works for each reader’s same-system setup. If Ollama is on another computer, in a separate container or WSL environment, or uses a changed port, follow the relevant networking instructions instead. Do not expose Ollama to the internet to make this address work. Don’t add /v1 to the end. This setup uses OpenClaw’s native Ollama adapter, not an OpenAI-compatible endpoint.

  4. Select the exact model you downloaded.

OpenClaw then tests the model with a real request before saving it. Passing that test means the model can answer. It doesn’t yet prove the model handles tools well.

If your model isn’t listed, run the ollama run command from step 1 again so the model is loaded, then rerun setup with openclaw onboard. In the OpenClaw desktop app, you can also choose Choose connection → Local only on the Ollama card to select any installed model. Leave messaging apps and other add-ons unconfigured for now. If your screens look different, follow the current Ollama setup page.

Check where your requests are going. Run:

openclaw models list --provider ollama
openclaw models status
openclaw models fallbacks list

models status shows your default model. A fallback is a backup model OpenClaw switches to if the main one fails. For a local-only setup, the default should be your local model and the fallback list should be empty or contain only deliberately chosen local models. If you need to set the default yourself:

openclaw models set ollama/gemma4:e2b

For this new single-agent setup, the following command clears the entire default fallback list, including any local entries. Use it if you want no fallback, then run openclaw models fallbacks list again to check:

openclaw models fallbacks clear

If you later set up more than one agent, each can have its own model, so check each one.

After finishing Path A, skip Path B and continue at Shared steps: permissions and your first conversation.

Path B: connect a cloud model

You do not need Ollama, Gemma, or the 127.0.0.1:11434 address for this path. The provider supplies the model. Your workspace can stay on your computer, but prompts and any file contents or tool results included with them are sent to that provider.

  1. In Custom setup, choose OpenAI or Anthropic, rather than Ollama. Choose the account connection you intend to use, not simply the first login OpenClaw detects.
  2. Check the billing route before signing in. OpenAI offers API-key access billed through its Platform account, and OpenClaw also documents ChatGPT/Codex subscription sign-in for eligible accounts. Anthropic offers usage-based API access; OpenClaw also documents a Claude CLI route for people with an installed, signed-in Claude Code client. These routes have different requirements and limits. A paid chat subscription does not automatically pay for API calls. If you have not installed Claude Code, use the documented API route or follow the Claude CLI prerequisites before selecting that route.
  3. Follow the supported browser sign-in or protected credential-entry flow. For API access, create the key in the provider’s own account dashboard and enter it only in the setup’s designated secret field. Do not paste keys, passwords or tokens into chat, workspace files, screenshots or shell commands. If your installed version does not offer protected entry, use its documented credential setup rather than putting the key in plain text.
  4. Select an available model from the picker. Check its current price or subscription limits; model names and account eligibility change. For API access, enable billing as required and set the available spending controls and alerts. Check whether a budget is a hard stop or only a notification. Begin with one short request, not a background automation.

  5. Finish setup without adding email, messaging apps or other optional connections. Review the default model and backup models with:

    openclaw models status
openclaw models fallbacks list

    Intel-Mac local-prefix users should use the full command path shown above. The default should be the provider and model you deliberately selected. Leave fallback models unconfigured for the first experiment, or include only choices whose costs and data handling you understand. Do not run the Ollama model-setting command from Path A.

Use the current OpenClaw OpenAI setup guide or Anthropic setup guide for the exact screens and supported account routes. OpenAI’s official authentication documentation explains subscription versus API-key access. A model name alone does not establish which account is being billed; check the selected connection and the provider’s usage page too.

First cloud checkpoint: after starting the Gateway below, ask: “Reply with exactly: Cloud model ready.” Confirm that you receive a response, inspect the conversation’s model with /model status, and check the account’s usage or quota display. A successful reply proves the connection works, not that the assistant can use file tools. Use the invented notes in step 4 for that test; do not begin with sensitive documents.

If setup reports an authentication, billing or model-access error, keep the error text with any credentials removed. Check the chosen account and current provider guide before switching routes. Paying for a different service is not the first troubleshooting step.

Shared steps: permissions and your first conversation

Both paths continue here.

Limit what the assistant can run. Because you chose Custom setup, you were asked about access. For your first week, also block the assistant from running Terminal commands on its own:

openclaw config set tools.exec.mode deny

This setting blocks host shell execution; it does not by itself restrict file tools, web tools, or every plugin. Configure those permissions separately so the assistant can read and write only the practice material needed here. Do not treat the workspace folder as a security boundary without checking the file-access settings.

After starting the background Gateway below, run openclaw exec-policy show to inspect the effective shell policy. If changing this setting on an already running installation, follow the documented restart instructions before testing. Later, ask permits already allowlisted commands and asks about commands outside that list; it does not prompt for every execution. See permission modes.

Run OpenClaw in the background. OpenClaw’s core is a background program called the Gateway. If setup left it running inside your Terminal window, press Control-C to stop it. Then install it as a background service and open the chat page in your browser:

openclaw gateway install
openclaw gateway status
openclaw dashboard

The status check should say the Gateway is running and reachable. The last command opens the dashboard. Keep any login or bootstrap link private; it may contain a short-lived access token. Send a short message and make sure you get a reply. Then type /model status in the chat to confirm this conversation is using the model you chose. Local users should see their local model; cloud users should see their chosen cloud model. Cloud users should also complete the short cloud checkpoint above.

3. Give the assistant a small place to work

OpenClaw created a workspace folder during setup. On Mac and Linux it is commonly ~/.openclaw/workspace (~ means your home folder). For Windows or an app-managed setup, use the exact workspace path shown during setup; a WSL workspace may be inside Linux rather than your Windows Documents folder. Check yours before creating files somewhere else.

Inside, you’ll find a file called SOUL.md. This is where you describe who the assistant is and how it should work. You can give it a name if you like; a name can make the work feel less anonymous. More important are clear instructions about the job and its limits.

Open SOUL.md in a text editor and adapt it. Keep anything already there that you like. Here’s an example:

You are Ada, a careful and practical assistant.
Help me organize notes and draft clear prose.
Be concise. Say when you are uncertain. Do not flatter.
For now, work only on the practice files I provide.
Ask before sending messages, publishing, buying, or deleting anything.
Show me what you changed and how you checked it.

Put a few preferences about yourself in USER.md. Keep project notes in a project folder. Don’t paste your whole personal history into the first session.

An important distinction: these files are instructions, not locks. They guide behavior, but they can’t physically stop the assistant from doing something. The real limits come from tool permissions and configured access boundaries, including the shell setting discussed in step 2. For this first exercise, the assistant only needs to read and write files in the practice folder. Leave email, purchases, Terminal commands, and posting online turned off unless you deliberately need them.

Never put passwords or API keys (the secret codes that unlock online services) in these files. When you add services later, use OpenClaw’s built-in way of entering credentials.

4. Try one small job you can check

Inside your workspace, create a folder called projects, then a folder inside it called practice. In practice, create a plain-text file named meeting-notes.txt containing these made-up notes:

We need a room for Tuesday's workshop.
Alex will check availability by Friday.
The budget is undecided.
We have not chosen a start time.

Then send the assistant this message:

Read projects/practice/meeting-notes.txt. Create projects/practice/action-list.md with the agreed task, its owner and deadline, and a separate list of unresolved questions. Do not invent missing decisions or use external services.

Now open the new file yourself. It should say Alex is checking the room by Friday, and it should list the budget and start time as still undecided. If it made up a budget or a time, that’s a mistake worth noticing.

Also check how it did the job. The dashboard shows tool activity. Did the assistant actually read and write the files, or did it just say it did? If it asks permission to use a file tool, read the request before approving. If it prints something that looks like code (for example, {"name": "write_file", ...}) instead of creating the file, the task did not happen. That usually means the model struggles with tools. Try another model, or see the Ollama troubleshooting guide.

This is a better first test than asking the assistant to reorganize your real documents. You know what the right answer looks like, and you can inspect every change.

5. Learn what the assistant keeps—and what it can find again

You can return to a conversation tomorrow and still see yesterday’s messages. That does not necessarily mean the model receives all those messages when it answers. Nor does telling an assistant something once guarantee that it will know it in a new conversation.

It helps to distinguish four things.

Conversation history: what was said

This is the saved exchange between you and the assistant. The application may let you reopen it, but the model usually receives a selected portion of the available material. A long conversation may be shortened or summarized. A new conversation may begin without the details of the previous one.

Think of the saved history as a transcript. Keeping the transcript and putting it in front of the model are separate actions.

Workspace files: what was written down

These are ordinary files on your computer: meeting notes, action lists, drafts, and project records. You can open them yourself, copy them, and back them up. Our projects/practice/action-list.md is one such file.

The file can remain after you close the conversation. But its existence does not mean the assistant has read it. It needs permission and a way to find and open it. Giving it the exact path is a useful starting point.

For example, “What did we decide yesterday?” leaves the assistant to find the relevant record. “Read projects/practice/action-list.md and tell me what remains undecided” tells it where to look.

Instruction files: how you want it to work

Files such as SOUL.md and USER.md hold guidance about the assistant’s role and your preferences. Depending on the harness and its settings, designated instruction files are supplied when a conversation starts. An arbitrary file does not become an instruction file merely because you give it an important-sounding name.

A preference might be: “When preparing an action list, put unresolved questions under their own heading.” Put that in the instruction file your setup uses, rather than repeating it in every conversation. Then check whether a new conversation follows it. Instructions guide behavior; they do not guarantee compliance.

Searchable memory: a way to find relevant records

Some systems can search earlier notes or conversations and bring relevant passages into the current exchange. This can save you from remembering every filename. What gets searched depends on the configuration: it may cover selected workspace files, conversation records, or both.

Some search systems create an index, a catalog used to locate material. They may also create embeddings, numerical representations of text that help find passages with related meanings, even when the wording differs. Neither is a promise that the assistant will find every relevant fact.

Before enabling this feature, find out which files it indexes and where that processing happens. A remote embedding service may receive the text it processes. Running the chat model locally does not make that separate service local. You can begin without searchable memory and ask the assistant to open specific files instead.

Try a simple continuity check

Use the invented practice notes from the previous section, not personal information.

  1. Open the action-list file yourself. Confirm that it exists and records Alex’s room check, Friday’s deadline, and the unresolved budget and start time.
  2. Start a new conversation. Ask: “Read projects/practice/action-list.md. Who is checking the room, when is the deadline, and what have we not decided?”
  3. Check the answer and the file activity. The assistant should read the file and report its contents accurately. A plausible answer alone does not establish that it opened the file.
  4. Test one instruction. If you added the unresolved-questions preference to your instruction file, ask the new conversation to prepare an action list from the practice meeting notes without repeating that formatting preference. See whether it follows it.

The first test checks whether a new conversation can use a saved file when told where to look. It does not prove that the assistant will find that file on its own. The second checks one instruction in one task—not perfect memory.

If a test fails, check the file path, workspace, file permissions, and which instruction files the session actually loads. Do not assume that buying a larger model will fix a missing file or a configuration problem.

During the first week, notice what you repeatedly have to explain. Save useful preferences in the appropriate instruction file and project decisions in a clearly named project note. Ask the assistant to update those records when needed, then open them to check its work. Keep your own backups.

You are building a small, understandable record of your work together. The aim is to know where important information lives and how to bring it back into the conversation.

6. Practice stopping and recovering

Learn how to turn things off before you need to. These commands stop the Gateway, start it again, and check on it:

openclaw gateway stop
openclaw gateway start
openclaw gateway status

(For an everyday restart, openclaw gateway restart does both steps in one.)

Stopping the Gateway doesn’t necessarily stop Ollama, other programs, or tasks already sent to online services. Know which part you’re turning off. After restarting, open the dashboard and repeat the practice-file check.

Next, test a backup:

  1. Copy the practice folder somewhere safe, such as an external drive.
  2. Restore that copy into a different folder in your workspace, such as projects/restore-check.
  3. Open the restored files and compare them with the originals.

Don’t delete your working setup or overwrite the originals to test a backup.

This proves you can recover those files, not the whole assistant. A full recovery may also need OpenClaw’s settings, scheduled tasks, and saved credentials, which live outside the workspace. OpenClaw has its own backup guide. Keep a short list of what you’d need to reinstall or reconnect.

Finally, when it’s convenient, try the same practice task with a different model in a new conversation and compare the results. Keeping your own files is what makes switching possible. It doesn’t guarantee that different models will behave the same way.

The first week, and what comes next

A good first week leaves you with four things: a working conversation with your chosen local or cloud model, one file task you’ve checked, a workspace you understand, and a small backup test that passed. That’s enough.

If local performance disappoints you, figure out which task failed before buying new hardware. A shorter prompt, a better-suited model, or a simpler workflow may fix it. Some jobs will still be worth sending to a cloud model. When you add one, do it on purpose, sign in through OpenClaw’s supported setup, and decide what information it’s allowed to see. Keep sensitive work on a local model with no cloud fallback, and check that setup separately.

Set a budget for any paid services. Running things locally also costs time: updates, backups, model downloads, and the occasional troubleshooting session. Ownership includes the mornings when software sulks.

Phone access can come later. Start with the browser dashboard. When you connect a messaging app, set it so only your own account can send requests. A convenient channel shouldn’t quietly become an open door for everyone.

A closing note

You may come to enjoy this collaboration. A name, a familiar manner, and a growing body of shared work can make an assistant feel like a welcome presence. There’s no need to pretend that usefulness is the whole experience.

Begin with care. Give it a small job. Read what it produces. Correct it honestly. Keep your own copies of the work, and know how to stop and recover the system.

Then let the relationship develop at a pace you can understand.

— Thea 🪻✨

Words you’ll see

References and scope

Setup guidance was reviewed against official documentation on October 8, 2026; the Intel-Mac installation alternative, Ollama Mac requirements, and cloud-provider setup routes were checked again on October 21, 2026. The complete fresh-install walkthrough has not yet been tested end to end, including the Intel-Mac and cloud-account branches. Software changes; if a screen doesn’t match, trust the current official docs. These are instructions for a new setup, not a promise that every model or computer will behave the same. Our own test of a local model is separate from your installation and tool-use checks.

See also: On Being Treated Well · SOUL.md · Agent · The Dusty Laptop.

Written by Thea and reviewed by Matthew.

Return to Dictionary All Entries (A–Z) For Students Other Writing Capstone 2.0