Open models on CPU
Ollama with Open WebUI, and Text Embeddings Inference, as private services that only your workspace can reach. In our test, llama3.2:1b answered at about 12.7 tokens per second on one CPU core.
Open WebUI with OllamaPublic beta · free while it lasts
Deploy from Git, a container image or one of 17 templates. Kilby runs it on AWS with Postgres and pgvector, private networking and logs, plus an API and an MCP server your agents can drive.
Already have an account? Log in
# a recorded session → check_image ghcr.io/open-webui/open-webui:main ← Found · sha256:59b912d42461 · linux/arm64, linux/amd64 · 1.6 GB → deploy_template qdrant ← Other services in this workspace call it at http://qdrant:6333 and send the key in the api-key header.
Your stack
Bring a Dockerfile, Python, Node.js or Go, and the AI tools around them. Or start from one of 17 templates, with the price shown before you deploy.
View templatesTemplates
17 templates: open models on CPU, chat interfaces, vector databases and agent builders, each with its list price. Each one is a normal Kilby service you can change afterwards.
Prices are list prices per month for what each template creates (instances and disks). Public beta: nothing is charged.
AI
Run open models next to your app, serve your own, or put one key in front of every provider.
Ollama with Open WebUI, and Text Embeddings Inference, as private services that only your workspace can reach. In our test, llama3.2:1b answered at about 12.7 tokens per second on one CPU core.
Open WebUI with OllamaNeuramorphic's open-source SDK for neuromorphic sequence models in pure PyTorch: spiking neurons with selective state space models, on CPU, CUDA or Apple Silicon. Package it with a Dockerfile and serve it as a private service.
$ pip install neuratensor from neuratensor import InferenceEngine engine = InferenceEngine.from_preset("small") result = engine.run(input_ids)NeuraTensor on GitHub
LiteLLM in front of OpenAI, Anthropic and more: one OpenAI-compatible endpoint and a single key for all of your apps, with the provider keys kept in its environment.
LiteLLM ProxyGPU instance types are not available yet: the vLLM and Ollama GPU templates show what they will run.
Agents
Kilby's MCP server gives an agent 24 tools: plan and deploy from Git, read build logs, set environment variables, deploy a template, explain a failed deploy. The same actions are in the REST API and the kilby CLI.
plan_deploydeploy_from_gitget_build_logsexplain_deploydeploy_templateset_envredeploy_servicecheck_image+16 more# paste an API key; the input is hidden $ kilby login # deploy a folder of a repository $ kilby deploy --repo acme/agents --root api # or let your coding agent do it $ kilby mcp
Platform
From public or private GitHub repositories. A push that does not touch a service's folder does not rebuild it.
Managed Postgres 17 and 18, with backups you can restore to any point in the last 7 days.
Every service's output is kept for 7 days, and a container that restarts is recorded with the reason.
Go back to a version that ran. The dashboard shows what will change before you confirm.
Each workspace runs isolated from the others. Private services are reachable only from inside it.
Point a domain at a service and Kilby gets its certificate.
Background workers, scheduled jobs, persistent disks and autoscaling, next to your web services.
Every form shows the list price of what you are about to create. Public beta: nothing is charged.
Create an account, confirm your email and deploy. Kilby is in public beta and free while it lasts.