Run a focused pilot.
Start with one team, a qualified model and your own hardware. Follow the source-build quick start, connect a test application and record the measurements that matter to your workload.
Start a pilotThe enterprise arm of TensorFold
StarLightEnterprise scale.
Open source, always.
Free, fully open-source AI infrastructure designed for large-scale enterprise deployments. Connect your models, GPU servers and teams with shared identity, deployment, data policy and usage controls.
- Free to use in full
- Self-hosted
- Apache-2.0
One platform. Every AI request.
Team applicationsDeveloper toolsAI agents
StarLightYour organisation's AI control plane
Company sign-inProject permissionsBudgets & routing
Your private fleet
Macs & GPU servers
TensorFold native inferenceApproved providers
Optional cloud routes
Same policies & budgetsOne endpoint for your teams. You decide which models are available and where requests can run.
Inside StarLight
Your AI estate.
One pane of glass.
See what is running, who is using it and where requests go. Manage your machines, access and budgets from the same console.
Screenshots of the running StarLight console with example organisation, hardware and usage data. Figures are illustrative.
Operations overview
Live throughput, running requests, connected users and fleet availability in one view.
Fleet management
Inspect hardware and replica health, enrol machines and drain or take nodes offline.
Project controls
Set data classification, approval rules, credential handling and project membership.
Budgets & limits
Cap concurrent work, tokens and spend, with temporary boosts when you need them.
Evaluate with your own team
A clear path to
your first deployment.
Choose a real workload, review the controls and agree what success looks like. Use the free software yourself, or arrange direct assistance through project backing.
Give reviewers the full picture.
Share the architecture, identity and data controls, operating responsibilities and open validation gates with your platform and security teams.
Read the evidence with its scope.
See the recorded control-plane tests, real-engine check and deployment checks. Identify which hardware, identity, recovery and security gates your rollout still needs to close.
Review deployment evidenceCalculate costs from your inputs.
Account for hardware, measured power, hosting, staff time, provider charges and project backing. Compare equivalent workloads using your own figures.
Open the cost plannerFrom individual machines to shared infrastructure
Give every team access.
Keep control of the fleet.
TensorFold runs the models. StarLight operates the service around them, from a person's first sign-in to the deployment handling their next request.
Company sign-in. Project-level access.
Connect your identity provider, map groups to roles and make the right models available to each project. Provision and deactivate users through SCIM.
- OIDC & SAML single sign-on
- Users, groups, roles & service accounts
- Per-project model grants and token caps
Deploy models across the machines you own.
Organise nodes into pools, place model replicas and manage deployment revisions. Keep models warm or start them on demand as work arrives.
- GPU-aware placement & node pools
- Warm, on-demand & batch deployments
- Rolling updates, drain, offline & revoke controls
Set the budget before the work starts.
Apply project limits across local inference and approved cloud models. Meter tokens and costs, cap concurrency and give interactive work priority over batch jobs.
- Request, token & spend limits
- Shared admission and priority queues
- Local chargeback & provider usage accounting
See what is running. Act from one console.
Follow live throughput, queued work, active users, node load and model availability. Manage the fleet and investigate usage without jumping between machines.
- Live operations console & usage reports
- Prometheus metrics & Grafana deployment assets
- Audited actions & verifiable audit-log integrity
Private by policy. Flexible by choice.
Decide where your
AI requests can go.
Use local models, approved external providers, or an offline installation. Access, classification and budget checks apply before a request is routed.
Local-only routing
Keep the request on your own machines.
Publish a local-only model alias. StarLight checks the caller's project, model grant, data classification and available budget, then routes to an eligible replica in your private fleet.
If no eligible local replica is available, StarLight refuses the request rather than sending it to a cloud provider.
01
Identify the caller
Company sign-in or scoped API key
02
Check project policy
Model grants, classification & limits
03
Illustrative workflow, not live deployment data.An eligible private replica
TensorFold on your managed fleet
Data classification
Control eligible models and routes for public, internal, confidential and restricted requests.
Credential guard
Detect credential patterns and block requests or restrict their routing under project policy.
Authenticated nodes
Enrol nodes over mutual TLS and authorise engine requests with scoped credentials.
Governed agent workflows
Plan with a frontier model.
Execute on your private fleet.
Give a planner the brief and context you choose to share. Review the resulting plan, approve its exact contents, then let the owner's agent execute against local models within a defined budget.
Task tokens are scoped to the approved executor model and deployment revisions. Delegated tasks can narrow their permissions, never expand them.
Explore plans & local execution- 01
Choose what to share
The planner receives the selected context under your project's policy and budget.
- 02
Approve the exact plan
A project administrator reviews and approves the plan by its recorded hash.
- 03
Execute within the limits
The owner's agent uses a local-only task token with bounded spend, concurrency and delegation.
TensorFold Connect
Bring your existing tools.
Sign in with your company.
Give people a local endpoint for their desktop applications, coding tools and SDKs. Connect signs them into StarLight and attributes each request to the person and app that sent it.
Connect & managed rolloutmacOS, Windows & Linux
Command line and menu-bar or system-tray access.
OpenAI & Anthropic-compatible clients
Chat, Messages, Responses and supported model endpoints.
Managed desktop rollout
Configuration profiles, Jamf, Intune, Group Policy and Linux configuration management.
Choose how you operate
Your environment.
Your deployment.
Run the StarLight control plane on your infrastructure, then enrol the machines that serve your models. Postgres is its stateful dependency; gateways can run behind a load balancer.
Docker Compose
Bring up StarLight and Postgres on one host. Add the optional development identity provider to explore sign-in and provisioning.
Compose guideKubernetes
Deploy with Helm, your own Postgres and optional ingress, autoscaling, monitoring and network-policy configuration.
Helm guideLinux service
Use systemd with TLS, managed credentials and an external database. Configure the fleet through the console or declarative files.
Linux guideStarLight control plane · StarLight Node on serving machines · TensorFold Connect for desktop access
View the source & installation docsSupport funded by project backing
Free to deploy.
Support through backing.
The whole StarLight project is free to use, including for enterprise deployments, and will remain fully open source under Apache-2.0. Every improvement we make, including work developed with enterprise users, will be published on GitHub. Self-managed deployments do not include direct support.
Back StarLight
Organisations fund direct support through donations, sponsorship or ongoing project backing. That backing also funds development and maintenance for everyone.
Back the projectDeployment assistance
Back the project and work with us on installation, integrations and workload validation. We agree the scope of assistance with your organisation.
Long-term support
Project backing can fund ongoing technical support, maintenance of agreed versions and integrations, troubleshooting and upgrade planning. We agree the scope and duration together.
Direct support is provided through agreed donations, sponsorship or ongoing backing of the project. We agree support scope, supported versions and duration with each organisation. All improvements remain open source.
Design partners welcome
Bring a real workload.
Help shape StarLight.
We're inviting organisations to evaluate StarLight on their own infrastructure and share technical feedback. Work through deployment, identity, workload quality and operations together. All improvements will be published on GitHub.
Explore a design partnershipWork with the project
Deploy freely.
Build with us.
Deploy and manage the free software yourself, or back StarLight through donations or sponsorship for deployment assistance and long-term support. Tell us about your organisation, your requirements and how you want to back the project.
Discuss backing & support enquiries@tensorfold.devEvaluate against your requirements.
The source includes tested controls and deployment assets. Real-fleet load, live identity tenants, hardware qualification and independent security review remain part of production readiness. The qualification record separates completed checks from open gates.