TrylleDocs
Automation

Automations

Run repository-aware coding agents from events, schedules, or manual dispatches with controlled bot permissions.

Trylle automations run Claude Code or Codex in an isolated sandbox when repository activity matches a trigger. Each automation combines a workspace, bot identity, repository scope, runtime, model, instructions, and permission boundary.

Automations use Trylle-hosted AI

Automation runs use temporary credentials issued through the Trylle AI gateway and consume the selected workspace's AI balance. They do not use the provider key configured for local try ask, try explain, or try draft commands.

How an automation fits together

PartPurpose
WorkspaceOwns the automation, eligible repositories, billing, and administrative access. It can be a personal workspace or an organization.
Bot identitySupplies the actor name, repository installation, and maximum permission scopes.
Repository scopeLimits the automation to all repositories owned by the workspace or an explicit subset.
TriggerStarts work from an issue, pull request, comment, CI, repository, bot, or manual event.
ScheduleOptionally adds an hourly, daily, weekly, or monthly UTC cadence.
Runtime and modelSelects Claude Code with an Anthropic model or Codex with an OpenAI model.
InstructionsTell the coding agent what to inspect, decide, change, and report.
RunRecords status, target context, model, duration, credits, errors, and logs.

At runtime, Trylle creates a short-lived bot session token. Its effective permissions are the intersection of the bot's scopes and any narrower scopes configured on the automation. A run tied to one repository is also token-scoped to that repository.

Create your first automation

Create a bot identity, install it on the target workspace, and wait until its provisioning state is Ready.

Open Settings > Automations. Choose New automation for a blank definition or open Templates to start with a prefilled workflow.

Select the workspace, bot, runtime, model, repositories, triggers, optional schedule, and permissions.

Write the instructions, create the automation, then use Run to test it against an explicit repository subject.

Inspect the run history and logs before relying on automatic events or a schedule.

Read Create an automation for field limits, trigger coverage, model choices, and permission behavior.

Trigger families

FamilyAvailable events
IssuesCreated, updated, closed
Pull requestsOpened, updated, merged, closed
BotReview requested, mentioned
CommentsCreated
CICompleted, succeeded, failed
RepositoryCreated, pushed
AutomationManual

Repository filters apply before a matching event starts a run. Schedules use five-field cron expressions evaluated in UTC; the dashboard builder offers hourly, daily, weekly, and monthly choices at quarter-hour times.

Operate existing automations from the CLI

List definitions for the active account or an explicit workspace:

try automation list
try automation list --account OWNER

Read recent runs and change whether new work can start:

try automation runs AUTOMATION_ID --limit 20
try automation disable AUTOMATION_ID
try automation enable AUTOMATION_ID

Disabling an automation prevents new automatic and manual dispatches. It does not cancel a run that is already active.

Current CLI boundary

Trylle CLI 1.1.4 can list, enable, and disable automations and list their runs. It cannot create, edit, delete, or manually run an automation, and it cannot read automation logs or cancel a run. Use the dashboard for supported authoring, manual dispatch, deletion, and log workflows.

Cost and data handling

Automation instructions and relevant repository context can cause source code, diffs, issues, pull requests, comments, and related metadata to be processed by the selected AI provider through Trylle's gateway. Real provider credentials remain server-side; the sandbox receives a temporary run-bound route to the gateway.

Runs draw from the owning workspace's Trylle AI balance. A run can be blocked when the account cannot reserve AI credit, and model work performed before a failure can still be billable. See AI usage for metering and spend controls.

Continue with

On this page