Connect your computer to ClawPod: from capability installation to your first task

Documented procedure, not an end-to-end tested tutorial. This article follows the public Node Host 0.2.1 workflow. Its documented final connection validator did not run on the Linux Agent inspected for this article, and its observation path uses test-fixture data rather than establishing a live remote connection. Installation completion is therefore not proof of connectivity: the Agent must obtain live evidence through an available runtime status tool, or report that verification is blocked. This limitation does not mean customer onboarding is unavailable.

1. Install the Node Host capability first

To ask ClawPod Agent to carry out a small task on your Mac or Windows 11 PC, start by preparing the Node Host capability on the Agent. A node is a computer connected so the Agent can request work on it. Installing the capability and connecting your computer are two separate steps.

Try this request in your Room. It is an example based on the official installation procedure, not a required command or an official quoted prompt.

Find clawpod-node-host at https://github.com/Wondermove-Inc/clawpod-capabilities, install it on my Agent, and verify it. Prepare the linked Skill and CLI Harness together. If it is already installed, check its version and readiness first. Do not start connecting my computer yet.

The Agent checks the package and compatibility against the canonical Registry and verifies the installation. If capability management is not available yet, the Registry’s official bootstrap procedure comes first. You do not need to choose package files or installation directories yourself: ask the Agent to explain what is ready. An existing installation conflict or unsupported runtime must be resolved before continuing.

Look for “the Node Host Skill and its execution tool are both ready and validated,” not just “the files have been downloaded.” This article covers Node Host 0.2.1’s self-service path. If your installed version differs, check the differences before following it. Do not assume this Registry workflow corresponds to a particular ClawPod Cloud installation button or Marketplace listing.

2. Tell the Agent which computer to connect

Once the capability is ready, continue in the same Room:

I want to connect my Mac as a ClawPod node. Please prepare the installation script for me to run.

For Windows, say “my Windows 11 PC” instead. The Agent identifies the operating system and prepares the appropriate script. You are not expected to find a connection address or fill in placeholders.

Keep the computer on and connected to the internet. You need to be able to open Terminal or PowerShell yourself. On a work-managed computer, first check that software installation and device connections are allowed. You handle any required installation permissions or operating-system confirmation prompts.

3. Run the supplied script on your computer

An Agent cannot automatically operate an unconnected computer just because you mentioned it in a conversation. In this workflow, you initiate the connection setup locally.

• Mac: run the Agent’s script in Terminal.
• Windows 11: run it in PowerShell.

The script prepares the components needed to run the node. It may include installation steps for Tailscale or Node.js if they are missing. Tailscale provides the network connection between the Agent and your computer. Before running the script, read the Agent’s explanation of what it will install and connect. This article does not provide a generic script to reuse across accounts. Use the script your own Agent generated for this connection.

Already using ClawPod on this computer? Ask the Agent to check existing runtime and node installations for conflicts before running the script. It installs a global runtime package, which can replace an existing installation in the same location.

If sign-in is needed, the Agent provides a link or the script displays the next action. Sign in yourself, in your browser, using the same Tailscale account used for ClawPod. If you do not know which account that is, check with the Agent or your organization administrator rather than creating an unrelated account. If the Agent’s own connection is not ready, you may receive its sign-in link first. You complete sign-in and authentication checks yourself.

An ACTION: message followed by a stop does not necessarily mean installation failed. Complete the requested installation or sign-in action, then run the same script again. If the next step is unclear, ask:

The script displayed an ACTION message. I will share only the message, without passwords or tokens. Please tell me the one thing I need to do next.

Some script messages may appear in Korean; share only the redacted ACTION: line and ask the Agent to explain the next step in English. Never share the full script, tokens, or raw logs.

4. Confirm the connection, then try one small task

When the script finishes, tell the Agent it has finished and ask it to verify this computer. A completion message in Terminal or PowerShell is not yet proof that the Agent can reach it. The documented workflow has the Agent watch for the matching node request and handle pairing; connection verification is a separate acceptance check, not a guaranteed result of that step. Pairing connects this computer as a node of the intended Agent. The 0.2.1 default workflow has the Agent handle the matching request; you do not need to assume there is a separate management screen where you must find a request number. Your runtime policy may require additional confirmation.

Ask for a report that distinguishes:

1. Which computer was connected.
2. Whether the live connection was verified.
3. Whether the task you want to request is currently available.

Ask the Agent to use an available live node-status tool to confirm the intended computer’s identity and current connected state. If no such tool or evidence is available, it must report “verification blocked” and stop before the first task; generated identifiers and test-fixture results do not count. This is an acceptance requirement, not a verified feature of the 0.2.1 Harness.

A connection does not automatically grant access to every file, browser, or screen-control function. Start with a small check that changes nothing:

On the computer we just connected, check only whether basic command execution is available. Do not read or change files or install programs. Tell me which check you ran and its result.

If execution requires approval, review the target and operation before proceeding. If your goal is screen control or browser work, check that capability separately after the basic connection succeeds.

5. Three useful tasks to try on your own computer

Once the live connection and command execution are verified, choose a task below. These are example requests and expected results, not tasks executed for this article. Node Host prepares the connection; separately available node-execution tools do the work. The Agent must use the named computer, not its Cloud workspace. The Guide’s /workspace filesystem access does not automatically cover your computer’s files.

All three examples are read-only and can apply to Mac or Windows 11 when the target permits the required commands. The Agent chooses commands for the detected OS and checks existing tools; no extra application is assumed installed. For file tasks, supply one exact path and begin with non-sensitive sample data—names and summaries may appear in the Room. Keep work within that target and path: no file changes, software installation, permission changes, or uploads to another service. If access or a suitable tool is unavailable, report the blocker instead of expanding the scope.

A. Check whether there is room for a large download

Planning to download a 10 GB archive? Start with a read-only disk-space check on the intended download volume.

On the computer we just connected, check the free space on the volume containing my Downloads folder. Tell me which computer and volume you checked, the free space, and whether it exceeds 10 GB. If the location is unclear, ask me which folder I mean.

Expected result: a capacity report, not a download. More than 10 GB free is only a space check; unpacking and temporary files may require extra room.

An illustrative response with invented values, not a measurement from this article:

• Computer / volume: my laptop / volume containing Downloads
• Free space: 42 GB at the reported check time
• Comparison: above 10 GB; unpacking space not assessed
• Changes made: none

Check success: the report identifies the right computer and volume, includes units and the check time, and explains the comparison. If surprising, compare it with the operating system’s storage information.

If it fails: “Was the connection unavailable, the disk query denied, or the download location unclear? Tell me the last successful check and the one detail needed next.”

B. Make an inventory before organizing a folder

Before tidying a project handoff folder, get an overview using only file names and metadata. Provide the exact folder path; contents need not be opened.

In the project handoff folder at the exact path I provide, list the immediate files with name, extension, size, and last-modified time. Summarize the count by extension and the five largest files. Do not open contents, enter subfolders, or follow links elsewhere.

Expected result: an inventory to help you plan a cleanup, not a reorganized folder. Similar names do not establish duplicate content.

Check success: confirm the target path, reconcile the total with the extension counts, and spot-check a few entries on your computer. Unreadable entries should be reported as skipped, not silently omitted.

If it fails: “Is the folder missing, unreadable, or too large for a bounded listing? Ask for a corrected path or suggest a smaller scope within this folder.”

C. Check a small CSV before using it in a report

Use a small, non-sensitive CSV export to catch missing values. This task reads the chosen file’s contents and needs a suitable CSV reader already available on the computer; it does not assume Python or Excel is installed.

Inspect only the sample CSV at the exact path I provide, using an existing CSV reader rather than splitting on commas. Report headers, data-row count excluding the header, and missing values per column, counting empty or whitespace-only fields as missing. Keep individual records out of the reply, and ask if the encoding, delimiter, or header is unclear.

Expected result: a compact data-quality summary, not a rewritten spreadsheet. If no suitable reader is available, the result should be a clear blocker rather than unreliable counts.

Check success: the report states the file, parser, encoding, delimiter, header assumption, and missing-value rule. Compare the counts against a tiny sample you can count yourself; quoted commas and embedded newlines must be parsed correctly, and malformed rows must not be silently skipped.

If it fails: “Is the blocker file access, a missing CSV reader, or the format? Tell me the smallest non-sensitive sample or format detail needed next.”

Routine read-only checks stay within the requested scope and the runtime’s execution policy. Moving or deleting files, installing software, uploading data, and changing accounts are separate actions requiring their own scope and confirmation. Browser and native-app work also need separate capability and permission checks; pairing alone does not enable them.

If setup stalls, do not start by reinstalling

A waiting state may simply mean the script or sign-in step has not finished. Ask the Agent for the last completed step and the one action needed now. When sharing an error, exclude passwords, tokens, and personal account details.

The connection is not complete yet. Check from the last successful step and tell me the one action I should take now. Do not start a duplicate installation first.

If provisioning is complete but the connection still fails, the Agent can continue with connection diagnosis. Stop and verify the target if an unfamiliar device appears or the intended computer is unclear.

Ask differently for a pause and a removal

To stop using the node temporarily:

I want to pause this computer’s ClawPod node connection. Tell me what will remain for reconnection and confirm the result of stopping it.

When you no longer need it, make the removal scope explicit:

Remove the ClawPod node from this computer. First explain what will be removed and what will remain, then verify that it no longer connects. Keep network tools that I use for other purposes.

Ordinary service removal and the installation script’s rollback are not the same operation. The 0.2.1 documentation says some connection settings remain after service removal; it also defines a script rollback path that removes the node service and package. Let the Agent determine which operation fits your request. Ask what actually stopped and what remains, rather than relying on “removal complete” alone.

Service removal is not the same as removing a potentially shared runtime package. Script rollback can remove that global package too. Ask the Agent to preserve shared runtime packages and dependencies by default and confirm any wider removal separately.

Scope of this guide

This article describes the self-service workflow defined in the public Node Host 0.2.1 documentation. We did not reproduce a complete installation on real Mac and Windows 11 machines for this article. Check your ClawPod Cloud environment’s policies, network access, installation permissions, and actual results with the target Agent. SSH-based provisioning and organization-wide network administration are outside this guide.

Start with one sentence: “Install the Node Host capability and check that it is ready.” Then identify your computer, follow the generated script and any sign-in steps, and confirm the connection with your first small task.