The model acts in your systems, within bounds you set
Agents that run on your fine-tuned models and your data. Give them your APIs as skills, your documents to answer from, and a place in the tools your teams already use. One access policy checks every action, and every change lands in a log with no delete path.
Your model, at work inside your systems
Six ways in. One access policy checks each of them, and one log keeps the record.
Acts in your systems
It calls your private APIs as skills, answers from your documents and works inside the tools your teams already use.
Within bounds you set
The organisation is tested first, a person is judged on their role and a key on its scopes, and a key never carries more authority than whoever minted it.
Logged with no delete path
Twenty-one actions land in an audit trail whose repository exposes no update and no delete, each attributed to the actor that caused it.
On your premises when policy requires
Nucleus delivers GPU hardware to your site, pre-loaded with your model, serving an endpoint on your own network.
Six ways your model reaches the work
One access policy governs every request, whether it arrives from a person or an API key. The organisation is tested first, so a request for another tenant's resource is refused before the policy looks at roles at all. When a key is turned away for its scopes, the refusal names the scope it needed.
Results from LLMs powered by Nucleus
delays resolved without a call
systems on one thread
incidents past their four-hour SLA
downtime prevented this month
What an agent did inside the systems it was given
47 trucks rerouted while they were still moving
Every incident past its four-hour SLA, escalated by name
Cable survey readings processed while the crew was still on site
What Deploy holds
Access
- One access policy for people and keys
- Four gates, the organisation first
- Three roles: owner, admin and member
- Four scopes: read, train, infer and admin
- A key's scopes capped by the role of the member who minted it
- A key bound to one project
- Revocation read on the next call
Evidence
- Twenty-one audited actions, each with actor, target, address and time
- Append only, with no update path and no delete path
- Keys are actors too
- Sign-in activity readable per address
- Data in ap-southeast-1 under Singapore jurisdiction
Ways in
- Tool definitions and tool choice on the Anthropic-compatible messages route
- Embeddings at /v1/embeddings and reranking at /v1/rerank, up to 256 inputs a call
- Retrieval files of query and passage pairs for fine-tuning a retriever
- MCP for editors and agents
- Signed webhooks
On your site
- GPU hardware sized to your workload, installed on your premises
- Your model pre-loaded
- An OpenAI-compatible endpoint on your LAN
- Air-gapped, with no runtime connection back to Nucleus
- Installed, monitored and updated by Nucleus
Six entries under Deploy
Harness
Permissions, scoped keys and audit
On-Premise Deployment
Our GPU servers, your site
Tool use
Your private APIs as skills
MCP
Your model in your tools
RAG
Answers from your documents
Webhooks
Signed events to your systems
Read on
Anthropic-compatible API
Serve Nucleus models to the Anthropic SDK and Claude Code with a base URL change.
Embeddings
Turn text into vectors for retrieval, and rerank a candidate set against a query.
Mandai Wildlife Group transforms animal care operations with Nucleus
Learn how Mandai Wildlife Group transformed their operations by integrating multiple systems and custom AI solutions into a unified platform.
Put your model to work inside your systems
Read how each way in works, or tell us what you want to build.