Build on Aktor

The builder: no code needed

Make, check and deploy a solution with forms. Download the package if you want it.

Solutions → New solution opens the builder. Edit in builder on an installed solution opens that version.

SectionWhat you set
BasicsTitle, name, version (with +patch/+minor), description
Data modelEntities and fields: type, required, indexed, enum values, which entity a ref points at. Renaming an entity updates everything that names it.
WorkspacesGoal, data access per entity, autonomy, connectors, agent, action and script stages (with "runs after", for each, calls per minute and run only if; scripts in a code editor), and triggers: a record change, a schedule, a webhook
AssistantInstructions, greeting, data access, limits, allowed sites, monthly cost cap, your own sign-in
Inputs & connectorsValues asked for at install; integrations picked from this server's plugins (secrets are asked for at deploy), or backed by one of your global connectors
Knowledge & skillsReference text agents search; skills in the Agent Skills format
EvalsExpected status, what the summary mentions, records written
Review & deployThe aktor.yaml it makes, the full check, downloads
  • Checked as you type: it uses the same validation the CLI and every deploy use. Each problem shows in its section, with a count in the left nav.
  • Deploy installs it from the page (or upgrades it, when that name is installed). It asks for inputs and secrets, then shows any webhook URLs it created. Nothing has to be downloaded.
  • Package (.zip) downloads aktor.yaml and its skill files, for git or the CLI. Edit YAML lets you change the text and apply it back to the forms.
  • Drafts are saved in your browser until you deploy or discard them.

Design with AI (bottom right) is a chat about the solution. Describe the whole app, or one change ("add a Vendor entity with name, phone and hourly rate", "add an eval for each workspace"). It can change every section. Each answer lists the changes, already checked against the server; nothing changes until you press Apply, and an applied change can be undone. It works on the draft as it is now, your own edits included, and asks when a decision is yours to make.

The designer receives field-validation rules and problems already present in the draft. If its proposal introduces errors, it can make up to two repair calls before showing the result, keeping the better attempt. Checked and fixed lists resolved problems from the first attempt. Remaining problems are still shown; repair calls count toward model usage and the organization's quota.

Type @ in the chat, or in any box agents read (goals and instructions, trigger instructions, eval inputs), to name a stage (@triage), a skill (@skill:handling-requests), knowledge (@knowledge:refund-policy) or a connector (@connector:sms). Each box offers only what its agents can use. Mentioning a connector connects it to that workspace, and renames carry over to mentions.

The CLI and SDKs, coming soon, will be alternatives for people who want them; anything made in one will open in the other.