A self-hosted AI assistant for restricted environments.
Curie is a self-hosted AI assistant and model gateway. It runs in your environment, and the only place it sends content is the model you configure.
Curie is built for restricted and air-gapped networks. Document reading, text recognition on scans, search, table queries and file export all run on the Curie host. Administrators decide who can use which model, at what cost, and Curie keeps an audit trail. Your own applications call models through the same gateway, under the same controls.
Who it is for
Organizations that work with controlled or sensitive documents and cannot send them to an outside service for processing. Curie is designed for these environments.
What it does
Documents
- Upload by file picker, drag and drop, paste, or a whole folder.
- Reads PDF, .docx, .doc, .pptx, .xlsx, CSV, text, .msg email files and images. Scanned pages go through text recognition on the Curie host.
- Answers cite the document and page each passage came from.
- Questions about a spreadsheet are answered by a SQL query that runs on the Curie host.
- A document library with shared knowledge bases, access grants and a 30-day trash.
- Export to .docx, PDF, .pptx, .xlsx and Markdown, rendered on the host, with your own style templates.
Governance
- Organizations and roles, with SAML single sign-on configured per organization.
- Features turned on per person and per organization.
- Usage limits and budgets, with token and cost reporting.
- An audit trail with a viewer, and forwarding to your syslog server.
- Controlled unclassified information marking, a classification banner and a use agreement.
- A ledger of API calls that records metadata and leaves out content.
Chat and workspaces
- Pick an agent, a saved set of instructions for a kind of work, and the model that answers. Either can change mid-conversation.
- Share a chat into a workspace, where team members read it and continue the conversation.
- Rate any answer, and send feedback to your administrators.
API and gateway
- An OpenAI-compatible API for chat completions and stateful responses.
- On models that support tool calling, tool definitions pass through to the model and tool calls return to your application.
- Personal keys, and service-account keys for systems.
- Requests from the API count against the same limits and budgets as chat.
Models and operations
- Connects to an OpenAI-compatible endpoint you run, or to your government-cloud model deployment.
- Liveness checks on each model backend, a built-in smoke test and a capability probe.
- A beta program that rolls new agents and backends out to a test group of users first.
- An onboarding walkthrough and a training center.
Applications on the gateway
Your own applications reach models through Curie. A new application starts with a Curie key and the address of your Curie host, and it runs under the controls your administrators already set.
Connect
An application written for an OpenAI-compatible API points its base URL at Curie and sends a Curie key. The model credentials stay on the Curie host, so no application stores them.
Tools
On models that support tool calling, tool definitions in a request pass through to the model, and the model's tool calls return to your application to run. An application that calls with a person's key can also turn on Curie's document tools, which search, read and query the documents that person can open.
Control
An administrator gives each system its own service account, with optional caps on requests, tokens and cost, and a key that can expire. Its calls count against your organization's budget, and Curie records each call without its content.
How it is delivered
Curie ships as a container image that you deploy in your own environment. It runs alongside a small retrieval service and a PostgreSQL database with vector search. An offline package covers networks with no internet connection. There is no hosted version.
Contact
Tell us about your environment and the work you want Curie to do. The sign-up form asks about your environment, the model provider you plan to use and the features you need.
Request access