Starlight Intelligence
Explore

Find your way through Starlight

Follow a completed mission, give your next project a brief, or build on the foundation.

Follow the workMissionsMake somethingMemory StudioMake it your ownFor buildersBuild togetherCommunity
Build once. Bring it with you.Capability FoundryFollow the thinkingResearchSee what is releasedOpen sourceQuestions and their evidenceStarlight AtlasJudgment that stays yoursTrust playbook2 to 144 perspectivesPlan a teamThe project and founderAbout StarlightYour first useful workflowStart with your agentTurn an idea into a crewMission composerMeet the crew and every seatAgentsCheck the prerequisitesInstall routerInspect the packagesDepartmentsPractice firstLearnGive an agent a jobAgent KitsSee what each agent contributedFollow a missionPick a working setTeamsSee the visual worldImagesFit to your workConfigureShared planePartnershipsA job in motionWorkflowsWho talks to whomAgent graphKeep the keyFoundationOne buildWorkshopsFour movesPathIdeas and evidenceTechnical notesContractsDevelopers

Starlight / Architecture workbench

Give your agents
a place to do good work.

A useful platform needs more than a clever answer. See where context, intelligence and human judgment meet—then carry a build plan into your own project.

01 / Choose the work

What would you like to make possible?

Design only · no inference cost
Choose a build pattern

02 / Follow the boundaries

Every layer has a job. And a limit.

workspace

Next.js · React · portable memory

Edit five context files and carry your brief between tools. Device recovery is optional, plaintext and browser-local.

Source: app/platform/studio.tsx

Explore this layer ↗

A prompt describes the job. It never grants identity, spending authority or tool permissions.

03 / Take the next step

Your build plan: shape a project.

This selection stays in the current tab. Download the plan to keep it.
Curator

Build in this order

  1. 01 · Prove the foundation

    Run the existing identity, budget, ownership and recovery tests. Review migrations and expiry operations before activation.

  2. 02 · Validate one real run

    Use a designated pilot account and an approved rate card. Verify provider usage, settlement and lost-response recovery end to end.

  3. 03 · Add capabilities deliberately

    Measure brief quality against a reviewed evaluation set before expanding access.

  4. 04 · Open a managed pilot

    Require tenant-isolation evidence, support and deletion procedures, payment-backed grants and observable spend limits before wider access.

What must be true before launch

  • A gated request cannot call a model before admission.
  • An interrupted request retains its run reference and cannot silently retry.
  • Use verified tenant identity, not tenant IDs supplied by the model.
  • Reserve and claim a bounded budget before inference; reconcile uncertain usage.
  • Recovery reads do not create a new model attempt.
  • Do not store prompts, memory or generated content in operational telemetry.