Model, profile and harness
Providers expose inference resources. A model performs a particular use; a profile organises those choices. A harness executes Tasks only and never receives the conversation round. Chat uses its internal controller and conversation resource.
Four text tiers and specialised resources
| Tier | Uses |
|---|---|
| Ultra low | Dream mechanisms using a model |
| Low | Dispatcher and conversation |
| Standard | Standard execution and Goal follow-up |
| High | High execution, Planner and Lab |
Reasoning effort is separate. Vision, documents, embeddings, images, transcription, audio/video analysis and media generation have dedicated selections. Voice is selected for an agent or user.
An explicitly selected profile is exclusive: missing resources are not silently borrowed from another profile. Without a specific selection, the agent or person follows the current profile.
A model for bounded decisions
The Decision resource supports specialised models such as TypeSafe’s Jev, selected in the OpenRouter preconfiguration. It complements the four text tiers with choices among defined options: Dispatcher selects an allowed route and effort, classification reuses a candidate topic or requests a new one, and memory retention chooses between ignoring, linking and extracting.
Galaris validates questions, returned choices and probabilities when supplied. The profile controls optional fallback to a text model for recoverable unavailability: the attempt is bounded, its reason is recorded and no confidence score is invented. The decision model guides the work; actions remain subject to Galaris permissions. Explore these mechanisms under the hood.
Refresh discovered model capabilities and filter the catalogue by resource type or document capability to select a resource suited to each use.
Connect and manage providers
A fresh installation includes OpenRouter, nine models and a preconfigured usage profile. Enter your API key to start talking to the Galaris assistant, then change selections, use another provider or remove the proposal. It does not impose a provider on existing installations. Unselected resources, such as particular voices or audio generation, are configured as needed.
For a supported personal ChatGPT connection, the subscription panel shows available usage windows, consumption and reset times. These limits apply to the provider account, not just calls through Galaris; unavailable information is identified.
Test a connection, discover models and select resources. Catalogue information can include modalities, context and available pricing; unknown costs do not become zero. Ollama also supports model management through its bridge.
| Family | Present integrations |
|---|---|
| Text and multimodal by resource | OpenAI, Anthropic, Gemini, OpenRouter, Mammouth AI, DeepSeek, Fireworks, Groq, Mistral |
| Other inference catalogues | Together, Cerebras, xAI, NVIDIA, Hugging Face, Cohere, Perplexity |
| Voice and media | ElevenLabs, Google Cloud TTS, Azure Speech, SunoAPI.org, BytePlus LAS |
| Local models and metadata | Ollama; models.dev for catalogue enrichment |
Actual accounts, models and modalities depend on the provider. Supported personal device-code authentication remains tied to its human owner.
The gateway adapts parameters to provider and endpoint contracts. An explicit rejection before generation can allow removing only the rejected optional setting and making a bounded retry. A timeout or partial stream does not trigger this removal; messages, tools, budgets and output format remain intact.
Choose each agent’s harness
| Harness | Role in Galaris |
|---|---|
| Internal Pydantic AI | Native execution, planning, deferred tools and effect checkpoints |
| Hermes | Autonomous per-agent runtime, shared memory and projected skills |
| Claude Agent | Per-agent SDK, Anthropic gateway and Galaris MCP |
| Codex | Managed runtime with tools and telemetry adapted to the common contract |
| DeepSeek Harness | DSH instance with adapted execution and streaming |
| Remote Chat Completions compatible | Configured server URL, model, authentication and capabilities |
Manage the lifecycle
Depending on the provider, inspect state, errors, logs and files; provision, start, stop or update the runtime. Long operations are tracked. Unfinished tasks can block an environment change.
Selecting a containerised harness does not automatically provision it. External providers currently admit one concurrent run per instance. Internal harness guarantees do not automatically extend to every adapter.
Architecture · API gateways · Compare approaches
Modes supported by the selected harness
The three modes are EXEC (standard execution), EXEC high and PLAN (planning at the high tier), depending on the harness’s capabilities.
| Harness | Available modes |
|---|---|
| Internal | Standard execution, high execution, high plan |
| Managed Hermès, Codex, Claude Agent, DeepSeek Harness | Standard and high execution |
| Generic compatible server | Standard execution |
The four managed runtimes currently use Galaris’s model gateway. A native SDK reasoning option alone does not enable high execution. The Dispatcher selects only permitted modes; a single choice needs no model routing call.
Administration sets policies and limits per provider. Effective capabilities combine implementation, configuration and target verification. A requested stop is distinct from a confirmed remote stop.
The personal ChatGPT connection requires recorded explicit consent from its owner. Revoking consent blocks calls even when authentication remains connected.
Native multimodal inputs in the internal engine
Attached images, audio, video and PDFs can be supplied directly to the model in chat and Tasks using the internal engine. Eligibility combines declared model capabilities, MIME type and what the transport can actually serialise.
Audio/video inputs use the same model through the compatible protocol; video requires a declared provider adapter. Support does not automatically extend to external harnesses. Access rights, size and a total budget are checked; refusals and unsupported formats are reported to the model. Media workflow details.
Give the model room to answer
For Chat/Responses calls without an explicit budget, Galaris uses the model’s published output capacity within the estimated remaining context. An explicit budget takes priority; when capacity is unknown, the provider default applies. Compaction and remotely stored Responses histories follow their own rules. An output cut short by the model limit does not by itself validate the Task.
Measure the differences before choosing
A model or harness change deserves evaluation on your own workloads. The Lab lets you compare models and mechanism configurations on the same cases: results, stability, time and available costs. Complement these measurements with a trial of the chosen harness using its own tools and capabilities.
Change engines while keeping connections in the hub
Hermes Agent, DSH (DeepSeek Harness), Codex, Claude and the internal harness can consume the agent’s authorised tools. Configure MCP servers and their functions in Galaris; the engine receives a projection appropriate to its permissions and execution context. Capabilities and recovery behaviour remain specific to each engine.
For example, Telegram can carry the request, Hermes can execute the task and a Notion MCP connection configured in Galaris can provide page operations. Choosing an engine does not dictate your chat interface or document workspace. See the setup and prerequisites.
Explore the ecosystem compatible with Galaris: inference providers, messaging, harnesses and services, with links to their reference websites.
Configure a provider and read its credit
After a successful connection test, provider configuration is saved automatically; subsequent changes to relevant fields are saved too. Keys remain masked: an empty input keeps the key, a new value replaces it and deletion requires confirmation. Removing a required key disables the provider.
Usage and credits shows known information for supported providers: ChatGPT, ElevenLabs, Mammouth AI, OpenRouter, DeepSeek and SunoAPI.org. The panel states the unit and account or key scope, refreshing on opening, on request and every five minutes while open. Unavailable values do not become zero; a balance is not a per-agent budget.
Document reading shared across engines
PDF, Office, text, HTML, Markdown, EPUB and images use common document preparation according to available formats and converters. For a large source, document_analyze starts private batch analysis; document_analysis_get reads coverage and results; document_analysis_cancel requests cancellation. The profile’s document model is used, falling back to standard text if absent. These analyses remain separate from business processes. Use documents.



