Automations
Automations are created in the Spore web app. You turn a linked node into a runner from the Nodes page. This is not a local JSON scheduler. An automation is a server-driven path: GitHub calls Spore, Spore assigns a runner, and that process runs the prompt you wrote.
The agent loop is Pi (the Pi coding-agent on the runner). Model turns go
through the runner’s Encrypted Local Proxy on
127.0.0.1:18790. The runner starts that proxy for the job. If the proxy
cannot run, the job fails and the prompt is not sent to the public API.
Interactive Cursor and OpenCode still use the same proxy for their own
sessions. Spore stays the control plane (claim, GitHub token, complete).
You do not put a GitHub token on the runner. Spore holds the GitHub App installation. The runner receives a short-lived token only while it is running a job, and that token is not stored.
The runner does not serve the model. Each model turn is a Spore chat completion through that proxy. Inference can run on the same machine or on another one.
Connect GitHub
Section titled “Connect GitHub”- Open Automations.
- Choose Connect GitHub and install the Spore GitHub App on the user or
organization that owns the repositories. On the GitHub install screen, pick
your personal account (for example
Chrisischris) or an organization — Spore does not pin the org. If GitHub only offers the Spore Intel org, an App owner must set Where can this GitHub App be installed? to Any account on the App settings page. - Spore records the installation on your account.
If the server has no GitHub App credentials, the page says so and automations stay off.
Create an automation
Section titled “Create an automation”On the same page, write an automation:
- Trigger: opened and ready for review, by default. New pushes
(
synchronize) and@sporeintel review/@spore reviewcomments are opt-in. On a matching comment Spore reacts with eyes (👀), then runs. A draft pull request does not run on open; it runs when it is marked ready, if that trigger is on. After a run that posts a pull request review,sporeintel[bot]appears under Reviewers — it is not in GitHub’s Request reviewers search. - Prompt: required. This is the instruction the agent follows. Use
Insert PR review starter for a suggested pull-request prompt. That
text is only a template — Spore does not rewrite prompts already saved on
an automation. Re-insert (and save) to pick up starter wording updates.
Spore does not add its own review rules. Prefer
create_pull_reviewso the bot shows under Reviewers. - GitHub account: the linked installation to use for this automation.
- Repositories: pick from repos that installation can access, or leave none selected for every repo it can see.
- Inference: My nodes (default) uses only models your linked nodes are already serving. Community prefers your node when it is serving the model, and a community node otherwise (that path spends credits). The runner machine stays yours either way.
- Model: an exact catalog id such as
qwen3-4b:8k. This is the chat model. The runner does not have to have it loaded. - Runner: Auto, or one linked node with Automation runner on. Auto uses an enabled node that is online and not already on an automation.
- Advanced: max turns (32) and timeout (900 seconds). CLI releases ship
a self-contained
libexec/automation-pi-runner(vendored Node + deps) — hosts do not need a separate Node install. Model turns go through the Encrypted Local Proxy on the runner (127.0.0.1:18790). The runner starts that proxy, and the job fails if it cannot. The runner calls Spore with linked-node credentials fromspore auth; a separatesk_spore_user API key is optional.
Runner assignment never falls through to another account’s nodes. If every eligible runner is already on an automation, the job stays pending and waits (same idea as inference when nodes are busy). If no enabled runner is online, or the pinned node has Automation runner off, is offline, or is not connected, the job is marked failed and the reason shows on the Automations page. A missing inference path is a separate failure, on a chat turn, not at runner assignment.
Enable a runner
Section titled “Enable a runner”On the Nodes page, open a linked node and turn on Automation runner. The switch is off until you enable it. Spore only assigns automation jobs to nodes that have it on.
A node already running spore start watches that setting and starts or stops
the automation agent. You do not start a second process for that common case.
The agent does not load the model. Each model turn is still a Spore chat
completion, on this machine or another.
A host that should run automations and not inference can use the optional CLI instead. The Nodes switch is still required. Spore will not assign jobs while it is off.
spore automation-runnerThe linked node credentials from spore auth are enough for those chat
completions. An optional sk_spore_ user API key still works if you already
have one. Keep some personal inference capacity online for the automation’s
model — this machine or another.
When a matching GitHub event arrives, the runner claims the job, reads your prompt plus the event (repository, pull request, commit, URL), and runs until it finishes, hits the turn cap, or times out. It then reports a short summary. Spore does not keep the full transcript.
The prompt decides what the agent posts. The PR review starter tells the agent to leave inline line comments on the diff, then a short review that says how many issues it found. It reads the diff, surrounding files, pull request comments, and any earlier review before it posts. The review ends with the model that ran the job.
Three review options sit on the automation:
- Autofix —
disabled(off),newBranch(push fixes to a new branch), orexistingBranch(commit onto the pull request branch, at most 3 times). - Incremental review — on reviews only the changes since the previous Spore review. Off reviews the full diff every run.
- PR summary —
comment,description, ordisabled. A summary includes a risk line.
That starter does not approve or request changes unless you change the prompt. Existing automations keep their saved prompt until you insert the updated starter and save again. The review options apply on the next run after you save them.
Pause or delete the automation in the web app to stop new jobs.
Related
Section titled “Related”- Running a Node
- Encrypted Local Proxy for Pi automation model turns, Cursor, and OpenCode
- Getting Started