A containerised stack you operate
The reference deployment separates frontend, backend, PostgreSQL with pgvector and optional services. Docker and declarative schema synchronisation support upgrades. The operator remains responsible for configuration, volumes, network exposure and recovery.
Start with a single API key
After a simple installation, an OpenRouter API key is enough to start talking. A fresh instance has a preconfigured profile and selected models; the first administrator is offered the Galaris assistant. Enter your key in Providers & Models, then open Chat. You can ask the assistant to guide you through Galaris using its documentation.
Use a key with access to the selected models; calls use your own OpenRouter account. You can change this configuration, choose another provider or use local models. Connect business services as you need them.
Local, cloud or hybrid inference: your choice
- Fully local: run models on your own machines, for example with Ollama, and connect Galaris to their addresses on your network. Inference can operate without a cloud model provider.
- Cloud: use configured providers such as OpenAI or OpenRouter with your accounts and keys. Galaris centralises access and call tracking.
- Hybrid: combine local and cloud models by profile and use case within the same instance.
For fully local infrastructure, also choose harnesses, tools and services compatible with that deployment, and configure the features you use for local resources. Each enabled external connection retains its own data flows. Models and harnesses explains these choices.
Self-hosted, local and free software are separate choices
Self-hosting Galaris means operating your own instance. The model generating responses may be local or remote: a cloud provider receives the data needed for its requests. Connected tools also have their own data flows. Review models and harnesses when preparing your configuration.
Galaris code is distributed under the CeCILL 2.1 licence. This licence covers the software; models and services retain their own terms. Follow the current GitHub installation guide, then prepare your team.
Map where data travels
Tasks, memory and documents retain durable state in the instance. Connected providers may receive selected input. Connection secrets are encrypted and resolved by the system. Check backups, retention, temporary files, exports and provider policies for the actual deployment.
Prepare for production
Test restoration, not just backup creation. Check network boundaries, secrets and denied operations. Define retention and deletion procedures, monitor health and incidents, and document upgrades. Sovereignty, compliance and availability depend on this concrete setup and its connected services.
Update without losing configuration
Before converting PostgreSQL enum types, DbAdmin checks permissions, value mappings, live state and dependencies. Unsupported cases are rejected before business data transformations. Supported conversions preserve constant defaults, type ownership, grants and comments. An error rolls back the enum conversion batch; this guarantee does not cover the entire upgrade, whose other phases have their own transactions.
DbAdmin derives the schema from models and synchronises constraints, indexes, privileges and reference data. Custom prompts can be compared with new defaults and replaced explicitly.
Backups must cover the database, files, executors and decryption material. A tested restore verifies what the mere presence of an archive cannot. Files held by an external provider remain under that provider’s responsibility.
Choosing a local model does not make remote voice, SaaS tools or external video generation local. Examine each enabled capability’s data flows. Tools and channels
Install and maintain an instance
The GitHub quick start guides installation. The installer prepares configuration and the services initialise the instance at startup. The same repository documents updates and backups; use it as the reference for your version. Synchronisation preserves your profile and assistant customisations.
Documentation and navigation maps are prepared during development with make docs-prepare, then kept with the source. make update includes this corpus without regenerating it, then refreshes and checks the shared index at startup. Authorised agents can search the text while vector indexing continues in the background.
If the embedding provider is overloaded or times out, vector indexing defers retries with increasing delays. Text search remains available while it waits.
