OmniCore-Local
A zero-trust, fully on-premise agentic gateway fusing MCP, OpenAI Swarm, NeMo Guardrails, and OpenTelemetry on a local LLM, no commercial APIs, no outbound calls, no data exfiltration.
Agentic systems usually mean handing data to commercial APIs.
Most agent stacks assume an OpenAI or Anthropic endpoint and outbound network access. For regulated or air-gapped environments that is a non-starter, the data cannot leave the building.
OmniCore-Local fuses Anthropic’s MCP, OpenAI’s Swarm, NVIDIA’s NeMo Guardrails, and Google’s OpenTelemetry, then powers all of it with a small open-source LLM (Llama-3-8B / Qwen-2.5-7B / Phi-3-mini) running locally on Ollama or vLLM. Zero egress.
A guarded request path with rails on both ends.
Each request flows Streamlit UI → FastAPI gateway → a hybrid input rail (regex deny-list, then a local LLM judge) → a stateless Swarm engine → an execution rail → an MCP stdio child over JSON-RPC 2.0 → an output rail → an OpenTelemetry flywheel. The same local `llama3:8b` serves both Swarm reasoning and judge classification.
The execution rail validates tool names and arguments against allow-lists and blocks directory traversal, sensitive paths, and shell metacharacters; the decoupled FastMCP server runs a second Pydantic + `shlex` validation pass before any subprocess exec.
Fail-open on the judge, deterministic repair for small models.
The LLM judge fails open on timeout, parse error, or endpoint outage so a flaky classifier never wedges the gateway. A deterministic JSON-repair pipeline rescues malformed tool calls from 7B/8B models (fenced blocks, single quotes, trailing commas), and a dedup loop guard terminates if the model re-issues an identical call.
Every span, input regex, input judge, swarm loop, execution rail, MCP tool call, output rail, run complete, is persisted to a traces file and streamed to a live audit matrix over WebSocket.
A complete agent platform that never phones home.
The output rail redacts leaked template brackets, tracebacks, and ChatML tokens before anything reaches the operator. 19 passing tests cover the rails and engine, and the whole system runs with zero commercial APIs and zero outbound calls.