Deploy your AI where the cloud can’t go.
Ghost runs your product on an open model inside your customer’s network. No internet.
Customers
6 on-prem sites. 4 on the latest release.
| Site | Version | Evals | Hardware | Last update | Status |
|---|---|---|---|---|---|
Regional bank Core data center | 2.4.02.3.0 | …23/24 | 2× A100 | Nov 6 | Updating to 2.4.0Rolled back |
State agency Sovereign cloud | 2.3.0 | 24/24 | 2× A100 | Sep 19 | Update ready |
Hospital network Outbound blocked | 2.4.0 | 24/24 | 2× L40S | Oct 28 | Running |
Defense supplier Air-gapped enclave | 2.4.0 | 24/24 | 4× H100 | Oct 30 | Running |
Electric utility Plant network | 2.4.0 | 24/24 | 1× L4 | Nov 1 | Running |
Insurance carrier Private VPC, no egress | 2.4.0 | 24/24 | 2× L40S | Nov 2 | Running |
$ ghost up acme-copilot-2.4.0.ghost currently running 2.3.0✓ vendor key e208441de2f697d1✓ signature and all 19 files verified✓ services running running 1 evals ✗ contract summary: answer does not contain "30 days"✗ evals failed; rolling back to 2.3.0✓ 2.3.0 is running again
For AI companies selling to
- Banks
- Hospitals
- Law firms
- Defense
- Utilities
- Government
- Insurers
How it works
- 1
Take it off the API
Swap OpenAI or Claude for an open model. Ship it as one signed file.
Building in your CI
acme-copilot 2.4.0
- OpenAI and Claude calls routed to a local open model
- App images from docker-compose.yml
- Model weights and eval suite
- Exact image IDs and model commits recorded
- Signed with vendor key e208441de2f697d1
Built acme-copilot 2.4.0 → acme-copilot-2.4.0.ghost (19 files, 96.1 MB)
- 2
Cross the gap
Their IT team imports it. Nothing connects back.
Your CI
Online
Regional bank
No internet
Air gap2.3.0-to-2.4.0.update.ghostFull release96.1 MBUpdate2.2 MBUpdate 2.3.0 → 2.4.0: 2.2 MB instead of 96.1 MB (2.3%).
- 3
Prove it, keep it running
Your evals run on their servers. A failed upgrade rolls back.
RunningRegional bank, on-prem admin
acme-copilot 2.4.0
Evals against your cloud baseline24 / 24- 09:12Update 2.4.0 imported by IT
- 09:14Signature verified, 19 files
- 09:15Evals passed, 24 of 24
Approve next updateExport support bundle
Building in your CI
acme-copilot 2.4.0
- OpenAI and Claude calls routed to a local open model
- App images from docker-compose.yml
- Model weights and eval suite
- Exact image IDs and model commits recorded
- Signed with vendor key e208441de2f697d1
Built acme-copilot 2.4.0 → acme-copilot-2.4.0.ghost (19 files, 96.1 MB)
Tested on real AI apps
Ghost demo 2.3.0 → 2.4.0
App and 2 models
Release 96.1 MB, update 2.2 MB. Update is 2.3%
n8n self-hosted AI starter kit 1.115.0 → 1.116.0
n8n, Ollama with llama3.2:1b, Qdrant, Postgres
Release 3.7 GB, update 268.2 MB. Update is 7.2%
Langfuse 4.75.0 → 4.76.0
7 services
Release 955.2 MB, update 389.5 MB. Update is 41%
Flowise 3.1.3 → 3.1.4
1 image
Release 1.5 GB, update 1.5 GB. Update is 97%
Three commands.
# once: your signing key; CI reads it from GHOST_KEY$ ghost keygenWrote ~/.ghost/vendor.key (keep private; in CI, set GHOST_KEY) and ~/.ghost/vendor.pub (give to customers).Key ID e208441de2f697d1# every release: reads ghost.yaml and your docker-compose file$ ghost buildhashing 19 filesBuilt acme-copilot 2.4.0 → acme-copilot-2.4.0.ghost (19 files, 96.1 MB)
Works with what you already run.
Ghost works with open models such as Llama, Qwen, Mistral, Gemma, DeepSeek and gpt-oss; runs on Docker Compose, Hugging Face, Ollama, NVIDIA GPUs and Linux with your evals and CI; and is tested with open-source AI apps including n8n, Langfuse, Flowise, open-webui, Dify, RAGFlow, LibreChat, AnythingLLM and Khoj.
From cloud API to their server.
- 01
Check
FreeWe find what breaks offline.
- 02
Convert
Fixed priceWe swap in an open model, tuned on your data.
- 03
Run
MonthlySigned updates. Automatic rollback if evals fail.
We stay on until it ships.
Every customer site, every update, with you.
Questions
Our product runs on OpenAI or Claude. Can it still run offline?
Yes. We swap in an open model tuned on your data.
Will an open model be as good?
Often close. We measure it on your tests first.
How is this different from Replicated or Zarf?
They’re built for Kubernetes. Ghost starts from your docker-compose file.
What hardware does my customer need?
Usually one Linux server with an NVIDIA GPU.
Got a deal stuck on on‑prem?
Send us your docker-compose file. Within a week, you’ll know what breaks offline.