The file exists. It is just nowhere useful.

The finished MP4 sits under /data/pi-agent/, inside a container your phone cannot open. Without an upload step: finish the video, open a computer, connect over SSH, find the output path, fetch it with scp, send it on. With rclone: finish the video, get a Drive notification on your phone, share it.

rclone is a file synchronization tool, the command-line counterpart to the Drive app, and it is already in the container: no apt install rclone, no separate version to manage. Every command has one shape:

Three reasons it suits this job, of which only the simplest direction — local file to Google Drive — is covered here:

• Mature. More than ten years of development; token refresh, retries and multipart uploads for large files are handled.
• Not locked to Google. More than 40 services: OneDrive, Dropbox, Box, iCloud Drive, Amazon S3, Alibaba Cloud OSS, SFTP, WebDAV. Either end can be cloud storage, so it will copy Drive to Dropbox or back S3 up to OneDrive.
• Direct. Files go from your own Home Assistant host to the provider you chose; Pi Agent does not relay them through a WoowTech cloud service.

Every other combination uses the same remote-and-path model.

Why rclone never asks for your Google password

Google no longer lets software sign in with your email address and password: anything holding it could reach the whole account — mail, Drive, your YouTube channel, payment services — and even software you trust can be compromised. It is the same delegated-access model as a service asking to read your Google Calendar.

flowchart TD
  A["rclone starts an authorization request
and prints a command to run
on a computer with a browser"] --> B["Google asks whether rclone
may access Drive, at the
scope you selected"]
  B --> C["Google issues rclone a token,
limited to that scope"]
  C --> D["rclone presents the token
for every later transfer"]
  D -.->|"you revoke it in
Google Account, Security,
third-party connections"| E["The stored token stops
authorizing rclone"]

Withdraw it any time from Google Account → Security → third-party connections. Type your Google password only on a genuine Google page whose address and account you have checked; close any other screen that asks for it.

This is the one chapter you type yourself

Most of this series happens by asking. Here you type: rclone config is interactive and runs inside the Pi Agent container.

What the agent can do on its ownFour tools in the left column; the one you select is documented on the right, down to its parameters.

• read, bash, edit, write is the whole list in the left column. bash is why the agent can help with terminal work — though the authorization below is still yours to do.
• read is selected, and the right-hand panel names its three parameters: path, required; offset and limit, optional. It reads a file you name, not whatever it likes.

1. Step 1

   Install Advanced SSH & Web Terminal

   Go to Settings → Add-ons → Add-on Store, search for Advanced SSH & Web Terminal, install it, open Web UI. No SSH client needed on your laptop.

2. Step 2

   You are not in Pi Agent yet

   The terminal opens in the SSH add-on's own environment: you may see system paths such as /config/ and /mnt/, and ls /data/pi-agent/ here may say “No such file or directory”. Correct, not a fault — that path is Pi Agent's persistent data.

3. Step 3

   Find your own container name

   Pi Agent's container is addon_<slug>, and a repository-installed slug usually starts with an eight-character repository hash, such as b9cf5676_woow_ha_pi_agent. The hash comes from the repository source, so different repository URLs give different names. Check Settings → Add-ons → Woow HA Pi Agent → Info for a Hostname or Slug; if yours shows none, do not infer it — discover it:

   docker ps --format '{{.Names}}' | grep pi_agent
4. Step 4

   Enter the container you just verified

   Substitute your own slug into docker exec -it addon_<your-slug> bash — for example docker exec -it addon_b9cf5676_woow_ha_pi_agent bash. Do not type the example unchanged unless it matches your installation exactly. A prompt like root@container-id:/# means you are inside; ls /data/pi-agent/ should now list sessions, models and rclone.

5. Step 5

   Stay there for everything that follows

   rclone config, rclone lsd and rclone config reconnect all need this container: the rclone executable and the persistent configuration live here. If you see rclone: command not found, stop and check the prompt and container name first.

Most prompts want the default. Two do not.

Run rclone config. Wording varies between releases; read each prompt before pressing Enter.

A finished wizard is not a working upload

Still inside the container, run rclone lsd gdrive:. It prints the directories at the root of your My Drive or Shared Drive:

A listing confirms read access. No output and no error is legitimate too, if that root holds no directories; create a folder and repeat for a clear answer. Three errors turn up often:

Now prove you can write. If your Drive root already holds a test.txt, use a unique name in both lines so the test cannot overwrite it:

Check Recent in the Drive app and read the contents of test.txt; only then delete it from Drive. Copying leaves the local file alone.

Point the pipeline at a folder

The pitch_video Skill needs a destination such as gdrive:PiAgent/Videos/, but which setting holds it varies by version: read the installed version's SKILL.md, README and configuration example, and never assume a field from another version exists. Three shapes are common, all taking the same value:

A small file job, finishedA different errand — here the agent wrote notes.md — but the same shape as reading a SKILL.md for you.

• Process details · 2 messages · 2 tool calls is collapsed by default. Open it when you want evidence of what ran rather than a summary.
• notes.md shows up twice: as a chip under the reply, and as a new entry in the EXPLORER list on the left. You are not taking the answer on trust.
• 187 in · 57 out · 1,152 cache R · $0.0032 is what it cost. Reading a SKILL.md is this size of job.

Create that Drive folder yourself rather than letting the copy make it, so you can check its location and sharing permissions first. Keep the trailing / in all three forms; with rclone copy, gdrive:PiAgent/Videos is a directory path too, and renaming a file on the way up needs copyto.

Now run the full pipeline from Part 11. Its last stage should log something close to this:

Then check the file in Drive — name, size, whether it plays. A success-looking log line is not verification. Do not delete the local output until the upload returned success and you have seen the remote file.

One file you should treat as a key

rclone writes the remote and its token to /data/pi-agent/rclone/rclone.conf:

The refresh_token fetches fresh short-lived access tokens without asking you again, which is what makes it long-lived. Anyone with a usable copy of this file may reach data inside the granted scope until you revoke the authorization.

/data/pi-agent/ is persistent add-on data, with two consequences. An upgrade replaces the add-on's packaged filesystem but leaves /data/ alone, so restarting or upgrading Pi Agent does not normally send you back through OAuth — see Part 15 first. And a Home Assistant backup can package that same /data/, rclone.conf included, when Pi Agent is selected and the backup succeeds. After restoring, run rclone lsd gdrive: and one harmless upload rather than assuming it worked.

Dropbox, OneDrive, S3 — or the NAS in your cupboard

Another destination is another run of rclone config: n for a new remote, a clear name such as dropbox, onedrive or s3, then the matching backend. Only the questions after the Storage step differ:

Test the exact remote before relying on it: rclone lsd dropbox:, then a harmless upload, as you did for Drive. Only then change the Skill's destination from gdrive: to dropbox:, onedrive: or s3:mybucket/.

You can keep several remotes at once: a Skill script can run rclone copy final.mp4 gdrive:PiAgent/Videos/ and rclone copy final.mp4 dropbox:PiAgent/Videos/. Each transfer spends its own provider's quota and can fail on its own, so check both exit statuses and both remote files before calling the run a success or deleting the local source. Two copies is redundancy, not a backup strategy — that needs retention rules and a tested recovery too.

The things that actually catch people out

Questions people ask at this point

Where to go from here