Runners in the cloud
Deploy to the cloud in the Connect a runner dialog runs Mob as a runner in your own account at a cloud host, on its free tier, so Mob stays on without your computer. Today the host is Render, set up in three steps. On Render’s free tier, conversations, memory and added providers reset on every restart or deploy; a paid instance with a disk keeps them (see Keeping conversations).
It is the encrypted runner with
no terminal: the relay and the trust boundary are the same, and Mob still
listens only on loopback inside its container. Every host builds the same
Dockerfile and reads the pairing JSON from MOB_RUNNER_PAIRING.
Render
Section titled “Render”- Pick Render and choose Create cloud pairing. It registers a runner and
shows its pairing JSON: the URL, runner ID, encryption key and the runner
credential itself
(
POST /api/runnerswith{"for":"deploy"}). There is no one-time code, because a free instance loses its files on every restart and could not keep what it claimed. The cloud stores only the credential’s hash. - Deploy to Render opens
render.com/deployfor this repository. Render reads the Blueprintrender.yamland asks forMOB_RUNNER_PAIRING; paste the pairing JSON there. - Render builds
Dockerfileand startsmob runner --serve. When the runner shows online in the dialog, choose Open runner.
The Blueprint is one free web service. Render may restart a free instance at
any time and its files do not survive that, so conversations, memory and
added providers reset on every restart or deploy; the runner reconnects by
itself from MOB_RUNNER_PAIRING. A free instance also sleeps after 15 minutes
without inbound requests. The runner’s relay connection is outbound and does
not count, so the image passes --keep-awake with Render’s
RENDER_EXTERNAL_URL: the runner requests its own public address every 10
minutes. One always-on instance uses about 744 of the workspace’s 750 free
hours a month. Deploys are manual (autoDeployTrigger: off), so a merge to
main never interrupts a turn; use Manual Deploy in Render to update.
Keeping conversations
Section titled “Keeping conversations”To keep conversations, upgrade the same service (about $7 a month plus $0.25 for the disk, billed by Render):
- On the Blueprint’s Settings page in Render, set Auto Sync to No, so
a later change to
render.yamldoes not switch the plan back to free. - On the service, change the instance type to Starter and add a 1 GB disk
mounted at
/data.HOMEis/data, so configuration,providers.jsonand the database then survive restarts.
Details
Section titled “Details”mob runner --servestarts Mob in the same process and exits when either stops, so the host restarts both together.--health-addranswers every request withokfor the host’s health check and serves nothing else.- With
MOB_RUNNER_PAIRINGset,mob runneruses those credentials instead ofrunner.json. The variable holds the encryption key and the credential, so it belongs in the host’s secret settings. To pair again, revoke the runner, create a new Render pairing, replace the variable and restart. - With no
providers.json, the runner uses the default free model, and on a free instance that is the only provider for now: Render’s free services have no Shell or SSH access, and keys are not read from the environment. On a paid instance with a disk, add providers from Render’s Shell withmob provider add … --key-stdin; they are kept on the disk. - There is no container runtime on Render, so the sandbox stays off.
- The model provider sees conversations, as on a computer, and so does Render, which runs the runner and holds its files and the pairing JSON.
Other hosts
Section titled “Other hosts”| Host | Free tier | Conversations | One-click |
|---|---|---|---|
| Render | Web service, 750 hours a month, no card; sleeps without inbound requests, which the runner prevents | Reset on every restart; a paid instance with a disk keeps them | render.yaml Blueprint |
Hosts checked on 2026-10-20 and left out: Koyeb (new accounts need a $29 plan
since the Mistral acquisition), Fly.io (no free allowance), Hugging Face
Spaces (Docker Spaces on free hardware need PRO since July 2026), ClawCloud
Run (shutting down), Google Cloud Run (scales to zero and needs billing, so
an always-on runner costs money). Railway and Northflank are candidates
(below). The dialog lists the hosts in
packages/app/src/hosts.ts; it shows a host picker once there is more than
one.
Candidate hosts
Section titled “Candidate hosts”Both have a free tier that suits a runner, but their one-click links come from a template made in the maintainer’s own account, so they are not in the dialog yet.
- Railway: $1 of usage a month, 0.5 GB RAM and a 0.5 GB volume, so
conversations would be kept. A template is made once in the dashboard
(GitHub source
fdkevin0/mob, an emptyMOB_RUNNER_PAIRING, a volume at/data); there is no documented API for creating or editing templates. After that, code updates need no work here: arailway.jsonwould carry the build and health-check settings, and Railway tells users who deployed a GitHub-based template when its root branch changes, and they choose when to apply it. Changing the template’s variables or volume stays manual. - Northflank: two always-on Sandbox services; it verifies a card at
sign-up, and free volumes are unconfirmed. Its template is JSON
(
apiVersion, aWorkflowofProject,CombinedServiceandSecretGroupnodes, arguments forMOB_RUNNER_PAIRING) that can sync both ways with a file in this repository, so the template’s configuration would be reviewed and updated through pull requests. Linking that sync and copying the share link are a one-time step in the maintainer’s account.