A shared foundation with multiple interfaces
The six levels of the agentic OS connect infrastructure, model providers, inference access, agent execution, shared services and applications. An external client can call the LLM gateway without adopting the Galaris interface; Nextcloud Talk, Matrix, OneBot and Telegram can provide access to your agents.
APIs and Janus expose the level you need. Configure Janus as a single front-facing model in a compatible client, then hand over to the selected agent. Channels and tools explains how to combine this foundation with business systems.
The harness executes a Task; Galaris owns its state
The conversation controller remains internal; only tasks use the harness assigned to the agent.
Galaris separates durable mission control from the execution engine. Agent identity, revisions, permissions, shared context, files and success criteria remain platform responsibilities. The chosen Task harness receives a bounded request.
Production invariants
Tools are resolved before execution. Access rules precede memory ranking. A file result needs an artifact and delivery evidence. A stream ends with one terminal result. Usage keeps its context. Internal checkpoints track effects; external harness guarantees require separate qualification.
Three execution timescales
The internal conversation controller handles the immediate exchange and can create a Task or Process. A Task uses a scheduler, leases, attempts and a selected harness. Voice has a separate low-latency path for speech and interruptions. The originating conversation is never handed to the Task harness.
Separate memory, files and context
Before a Task run, Galaris assembles canonical history, authorised memories and a versioned Working Set. The conversation controller builds its own context. Business objects remain authoritative in their domains; memory is not a replacement database. Canonical resource URIs avoid exposing arbitrary host paths.
Goals, Processes and evaluation
A Goal advances through ordinary Tasks and a separate evaluation step. Human questions create durable waits. Long external Processes keep their own run identity and events. Laboratory evaluates versioned cases and distinguishes a weak answer, candidate failure and judge failure.
Operate the whole system
The reference stack uses Docker, PostgreSQL and pgvector. Health, logs, incidents and correlation identifiers support diagnosis across model calls, tools and outputs. Harness supervision exposes lifecycle actions while each bridge retains its native protocol and configuration.
Inspect the contracts that matter
Start with architecture, then models and harnesses, APIs, Laboratory and operations. Deployment and data covers the boundaries to assess before production.
Qualify the deployed version
Harness capabilities and validation levels differ. Check the actual code version, driver contracts, enabled modules and instance configuration. Backups, restoration, load and operating commitments need verification on the deployment being evaluated.
Evaluate the platform at its boundaries
Models can be served to external clients. Agents can also be exposed as models, with Task-backed execution. An agent’s MCP endpoint exposes authorised capabilities. These three interfaces have distinct responsibilities.
Files retain their original provider, processes their engine, and Galaris its coordination objects. This separation makes access, costs and recovery inspectable at each boundary.
Tool inventory · Compare approaches
Durable inferences add frozen requests, attempts, replayable streams and control commands to the model gateway. This layer remains separate from Tasks, tool effects and file delivery.
Document applications and native media
The browser SDK galaris.datasets connects HTML pages to JSON data under access checks and versioned consent. It is separate from the 131 native MCP functions. Native multimodal inputs in the internal engine combine model and transport capabilities. APIs and SDK · Models and engines.