MCPlexor – MCP multiplexer that cuts agent context usage by 95%
I built MCPlexor to solve a token waste problem I kept running into with MCP-based agents. The Problem: MCP (Model Context Protocol) is great for giving LLMs access to external tools. But if you connect multiple servers (GitHub, Linear, Postgres, Slack), you end up with 40-50k tokens of tool definitions injected into every request – before the agent even does anything. On a 200k context model, that's 25% gone. On smaller models, it's worse. And most runs only use 1-2 tools. The Solution: MCPlexor sits between your agent and your MCP servers. Instead of loading all tool definitions upfront:…
What it does
In the maker’s words, at launch
I built MCPlexor to solve a token waste problem I kept running into with MCP-based agents. The Problem: MCP (Model Context Protocol) is great for giving LLMs access to external tools. But if you connect multiple servers (GitHub, Linear, Postgres, Slack), you end up with 40-50k tokens of tool definitions injected into every request – before the agent even does anything. On a 200k context model, that's 25% gone. On smaller models, it's worse. And most runs only use 1-2 tools. The Solution: MCPlexor sits between your agent and your MCP servers. Instead of loading all tool definitions upfront: Agent asks for a capability ("create an issue") MCPlexor routes to the right server using semantic matching Only relevant tools get exposed Result: ~500 tokens overhead instead of ~20k. Technical Details: - Written in Go, single binary, no runtime deps - Supports both stdio and HTTP transports - Stores credentials in OS keychain (macOS Keychain, Windows Credential Manager, Linux keyring via zalando/go-keyring) - Works with Claude Desktop, Cursor, Augment Code, or any MCP-compatible client The routing logic can run entirely locally using Ollama. No API calls, no cost, works offline. Business Model (for transparency): - Local/Ollama users: completely free - Cloud tier (on waitlist): we run inference on efficient small models instead of Opus/Pro, pass on savings, take a margin The bet is that routing is a narrow enough task that a fine-tuned 7B model does it as well as a frontier model. Early testing suggests this works. CLI uploads: https://github.com/arustagi101/mcplexor Install: `curl -fsSL https://mcplexor.com/install.sh | bash` Happy to answer questions about the architecture, the routing approach, or anything else.
Does the same job
all alternatives →
- MAMCPfinder – An MCP server that finds and installs other MCP serversApr 2026 · mcpfinder.dev · ▲9
I’ve been building and using agents heavily lately. The Model Context Protocol ecosystem is growing insanely fast, but discovering and configuring new tools is still highly manual. Every time I needed to connect an agent to a new service, I had to browse registries, figure out the transport type, identify required env vars, and manually update "mcp.json" files. So I built MCPfinder. It aggregates servers from the official MCP registry, Glama, and Smithery (around 25,000 combined entries) into a deduplicated, ranked catalog. But the real twist is the DX: MCPfinder is itself an MCP server :D…
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
