OpenClaude MCP Server

A high-performance control plane for Ollama-based local LLM sessions and background memory consolidation. Optimized for RTX 4090 environments and large-context model orchestration.

🚀 What is this?

If you need a reliable way to manage local LLM sessions, monitor VRAM usage, and maintain a structured memory between tasks without manual overhead—this is for you.

• Background Memory (KAIROS): Automatically consolidates session data into a persistent MEMORY.md using asynchronous background cycles.
• Ollama Control Plane: Comprehensive toolset for model management, health checks, and high-speed local inference.
• Hybrid Planning (ULTRAPLAN): Optional handoff to Claude Opus for complex reasoning, while maintaining 100% local tool execution.
• Fleet Observability: Integrated React dashboard for real-time monitoring of sessions, models, and memory status.

---

Quick Start

# Install dependencies
.\setup.ps1

# Start the MCP server + webapp
.\start.ps1

• SSE Endpoint: http://localhost:10932/sse
• Fleet Dashboard: http://localhost:10933
• Health API: http://localhost:10932/api/health

---

Available Tools

Tool	Action
start_session	Initialize a new Ollama session
send_prompt	Execute a prompt in an active session
kairos_enable	Activate background memory consolidation
kairos_disable	Halt background memory consolidation
list_models	Inventory of available Ollama models
model_status	Check VRAM and load status
ultraplan	Hybrid cloud/local planning cycle
fleet_status	Global health of the instance

---

Architecture

OpenClaude operates as a unified REST bridge for FastMCP, providing a robust integration layer between local LLM runtimes and agentic workflows. It is hardened for Windows environments with UV dependency management and subprocess isolation.

• Memory Management — KAIROS implementation details
• Hybrid Planning — ULTRAPLAN handoff logic
• Reliability & Hardening — Process management and stability

Prerequisites

• Ollama (running locally)
• Node.js (v20+)
• Python 3.13+ with uv
• Local model pulled (e.g., gemma2, llama3.3)

🛡️ Industrial Quality Stack

This project adheres to SOTA 14.1 industrial standards for high-fidelity agentic orchestration:

• Python (Core): Ruff for linting and formatting. Zero-tolerance for print statements in core handlers (T201).
• Webapp (UI): Biome for sub-millisecond linting. Strict noConsoleLog enforcement.
• Protocol Compliance: Hardened stdout/stderr isolation to ensure crash-resistant JSON-RPC communication.
• Automation: Justfile recipes for all fleet operations (just lint, just fix, just dev).
• Security: Automated audits via bandit and safety.

License

MIT