Skip to main content
EVOKORE// SKILLS / offline-mode
Agent-33markdownsha EE4A6C363DCE

offline-mode

documentation

Offline Mode

AGENT-33 is designed to operate fully within air-gapped or connectivity-limited environments.

What Works Offline

CapabilityOffline Behavior
LLM InferenceOllama with locally-pulled models — full functionality
Prompt OptimizationAPO loop runs entirely locally
Workflow ExecutionAll DAG execution is local
Memory / RAGPostgreSQL + pgvector — full functionality
EmbeddingsFalls back from Jina to Ollama embeddings automatically
Tool ExecutionShell, file, and local tools — full functionality

What Degrades Gracefully

CapabilityDegraded Behavior
Web SearchSearXNG returns no results → system uses local knowledge base
Web ReadingJina reader unavailable → URLs queued for later processing
Cloud LLMOpenAI unavailable → all routing goes to Ollama
Repo IntakeCannot clone remote repos → accepts local paths only

Queuing Improvements

When offline, the system continues generating improvement proposals but queues those that require:

  • External validation (web search for current data)
  • Cloud model evaluation (when local models lack capability)
  • Remote repo access

Queued items are stored in engine memory with status pending_connectivity. When connectivity returns, run agent33 sync to process the queue.

Data Sovereignty

All observations, analyses, and improvements are stored locally:

  • Engine database (PostgreSQL)
  • Local filesystem (docs/, engine/data/)
  • No telemetry or external data transmission

The offline_mode: true config setting enforces this by disabling all outbound network calls from the engine.

offline-mode — Evokore Skills