Orchestrator for
decentralizedself-hosted
AI agents

Advanced agents on top of local models

Each agent is an autonomous computational module with encapsulated state, a specialized system prompt and an isolated tool space. The Archangel is an HTN planner: it builds a graph of sub-tasks with explicit dependencies, allocates agents by specialization and coordinates them through a shared state bus.

Higher efficiency from local models

Three orthogonal mechanisms: (i) the context window is managed via semantic retrieval — at inference time a weighted sample by relevance score is supplied; (ii) parallel execution of DAG nodes with synchronization at dependency barriers; (iii) adaptive task routing by latency/quality/VRAM profile.

Unlimited context — the agent always remembers

Four memory layers: working (inference window), episodic (extractive summarization), semantic store (vector DB, ANN, cosine similarity with temporal decay) and procedural (behavioral patterns). Retrieval is a ranking task: k relevant fragments within a token budget.

Productivity monitoring for local models

Real-time observability stack: task completion rate, confidence calibration score, tool invocation success ratio, context utilization efficiency, VRAM·time. Hallucination detection via cross-validation between agents and consistency checking on factual claims.

Live behavioral correction algorithm

Behavioral monitor as a closed control loop: it computes topic drift score, uncertainty expression rate and task alignment index. On threshold breach: runtime modification of the system prompt, rebuilding the retrieval strategy and changing tool access scope. On a reasoning loop — a forced context reset.

100% offline — full data privacy

The entire inference stack, retrieval, tool calls and inter-agent coordination are confined to localhost with no outbound connections. Telemetry is architecturally absent. GDPR, HIPAA and SOC 2 compliance are structural consequences of the isolated runtime.

Agent orchestrator for complex multi-level projects

The scheduler builds a DAG with dependencies, priorities and resource constraints. Agents execute nodes in parallel; artifacts are passed through typed interfaces with schema verification. Online replanning: only the affected sub-tree is rebuilt, not the entire pipeline.

Integration with 40+ services and systems

A unified tool-calling interface on top of REST, GraphQL and WebSocket adapters. Each connector encapsulates authorization, rate-limit handling and retry with exponential backoff. New integrations are added without modifying the orchestrator core.

Built-in unlimited image generator

Local inference on Stable Diffusion / FLUX. Integrated as a first-class tool callable by an agent programmatically. VRAM management during parallel LLM inference and generation via dynamic offloading of weights between GPU and CPU.

Full voice communication — agents hear and speak

STT uses Whisper-compatible architectures with local inference; TTS uses neural vocoder models with low latency. Voice input is deserialized into a structured intent. Response synthesis is asynchronous — audio is streamed as it is generated.

Vision for any model — even small ones

A cascaded multi-model pipeline: small models analyze image structure and the results are aggregated into a structured text description for a text-only LLM agent. Vision capabilities without a multimodal LLM in the main stack — substantially lowers the VRAM threshold.

Large library of local models with specialization search

A catalog with metadata: benchmarks (MMLU, HumanEval, GSM8K, MATH) per task, VRAM at different quantization levels (Q4_K_M, Q5_K_M, Q8_0, F16), throughput on reference hardware. Search matches the task profile to model characteristics within resource constraints.

Full cloud-model support — Gemini, Anthropic, OpenAI, Grok, DeepSeek

A unified LLM abstraction layer: a single interface for local and cloud providers. The orchestrator routes tasks by a multi-criteria function: cost, latency, capability profiles and privacy constraints. Routing is configured declaratively.