Show and tell: omnibridge-mcp — Dynamic multi-server routing & protocol translation for MCP web clients

[johannschiffer]

Disclosure: Maintainer of omnibridge-mcp (MIT, available on PyPI). Posting in General to share an architecture pattern for aggregating and multiplexing Model Context Protocol (MCP) toolchains in multi-user web chat clients like LibreChat.

---

The Architecture Challenge: MCP Tool Overload in Web Clients

As web chat platforms (like LibreChat) expand MCP support, connecting multiple upstream MCP servers (e.g., database tools, scrapers, shell automation, cloud APIs) introduces two distinct operational bottlenecks:

1. Prompt Context Saturation & Attention Drift:
   Injecting dozens of raw MCP tool definitions directly into the system prompt exhausts context windows and degrades tool-calling accuracy. Models can struggle to distinguish between overlapping tool parameters when 30+ schemas are declared simultaneously.

2. Cross-Server Notification & Event Bus Isolation:
   Background asynchronous alerts (e.g. build completions, webhook receipts, database alerts) need a persistent, non-blocking queue that can dispatch events to the chat UI without blocking the active agent reasoning loop or losing messages during client reconnects.

3. Stdio Protocol Framing:
   In subprocess transports, unhandled diagnostic outputs to sys.stdout corrupt the JSON-RPC stream, causing host clients to drop the connection.

---

What omnibridge-mcp Does

omnibridge-mcp is a lightweight, zero-latency Model Context Protocol server that acts as a unified notification bus and multi-channel webhook router for AI agents.

1. Resilient Event Queuing with SQLite WAL

• Incorporates a high-throughput, embedded SQLite WAL (Write-Ahead Logging) queue.
• Incoming agent webhooks and notifications are written with ACID guarantees in sub-millisecond execution times.
• Agents can publish, poll, or drain events without locking active conversation threads.

2. Channel Multiplexing & Priority Filtering

• Exposes structured event routing: agents can categorize notifications by topic, severity (info, warning, critical), and channel destination.
• Reduces LLM token bloat by summarizing queued event streams into compact, structured payloads instead of flooding chat histories with raw event dumps.

3. Strict Stdio Protection

All connection heartbeats, database WAL checkpoints, and diagnostics route exclusively to sys.stderr, preserving 100% protocol integrity for host agent transports.

---

Usage & LibreChat Integration

Direct Execution via uvx (No installation needed)

uvx omnibridge-mcp

MCP Client / LibreChat Configuration (librechat.yaml)

mcpServers:
  omnibridge:
    type: stdio
    command: uvx
    args:
      - omnibridge-mcp
    env:
      OMNIBRIDGE_DB_PATH: "~/.config/librechat/omnibridge.db"

Exposed Agent Tools

• publish_event: Enqueues an alert or webhook payload with priority and metadata.
• peek_queue: Inspects unacknowledged events matching specific channel filters.
• acknowledge_event: Acknowledges and dismisses processed notifications.
• queue_stats: Retrieves real-time queue health, pending count, and throughput stats.

---

Links & Code

• PyPI Package: pypi.org/project/omnibridge-mcp
• Source Code: github.com/johannschiffer/omnibridge-mcp (MIT License)

Feedback from the LibreChat community on multi-tenant MCP event routing or webhook triggers is very welcome!