Toolport: Fix the tool-list bloat slowing your AI agent

Tool-list Bloat

The core problem Toolport addresses is the significant token overhead experienced by AI agents when integrating multiple tools, particularly those connected via MCP (Multi-Client Protocol) servers. As the maker, Tyler, explains, each MCP server typically dumps its entire tool list into the agent's context on every single request. This leads to a phenomenon dubbed "tool-list bloat," where a substantial portion of the agent's context window is consumed by tool definitions before any actual task processing begins. For instance, just three servers could cost approximately 24,000 tokens in definitions alone, severely impacting efficiency and increasing operational costs.

This inefficiency scales directly with the number of tools and servers connected, making it challenging for agents to handle complex tasks requiring a broad array of functionalities without becoming prohibitively expensive or slow. The constant re-sending of full catalogs on every turn, even if only a few tools are needed, represents a major bottleneck in the current architecture of many AI agent setups. The problem is not the AI model itself, but rather the inefficient management of the tools it needs to access.

Meta-tools

Toolport's innovative solution to this bloat is to position itself as a local gateway between AI agents and their MCP servers. Instead of directly exposing every single tool definition to the agent, Toolport presents a handful of "meta-tools." These meta-tools act as intelligent proxies, allowing the agent to search for specific functionalities only when they are actually needed, on demand. This lazy-loading approach drastically reduces the amount of information that needs to be held in the agent's active context at any given time.

The impact of this approach is substantial, with measured results showing up to a 97% reduction in tool overhead per request and approximately 90% fewer tokens consumed, all while maintaining the same task success rate. For a large catalog of 415 tools, the definition overhead dropped by an impressive 99.5%. This efficiency gain is critical for scaling AI agent capabilities, enabling them to work with a much larger and more diverse set of tools without being bogged down by the associated token costs and performance penalties.

OS Keychain

Beyond efficiency, Toolport also integrates robust security and privacy features, leveraging its position as an intermediary gateway. A key aspect is its handling of sensitive information: secrets and API keys are stored securely in the operating system's keychain, rather than being embedded in client configurations as plain text. This significantly enhances security by preventing credentials from being exposed or easily compromised, a common vulnerability in multi-client setups.

Furthermore, Toolport actively monitors for potential threats such as "rug-pulls" and "tool poisoning," flagging suspicious activities to the user. It also distinguishes between injection-looking tool output and legitimate data, treating the former as data rather than instructions to prevent malicious code execution. For potentially destructive calls, Toolport can pause execution and require explicit user approval from the application, providing an essential layer of human oversight and control over agent actions. The project is also free, open-source, and MIT licensed, operating entirely locally without routing data through external servers, reinforcing its commitment to user privacy and control.