Agent Frameworks Are All Competing on Tool Calling—VCPToolBox Is Betting on Memory and Permissions First

VCPToolBox is an Agent infrastructure project that Dan Kornas recommended in a tweet. It is hosted on GitHub under the CC BY-NC-SA 4.0 license: attribution, non-commercial, share-alike. The problem it solves is very specific: every Agent has to rebuild memory and reconnect tools from scratch, which is too much repeated work. VCPToolBox's approach is to add a system layer between the model API and the frontend, bringing plugin execution, memory processing, and model routing into it.

The feature list is substantial:

• Text-based tool calling: uses a marker protocol and does not rely on the model's native function calling.
• Distributed plugins: synchronous, asynchronous, and other plugins can run across servers.
• Associative memory: surfaces relevant memories and folds low-relevance context into summaries.
• Model routing: selects models based on conversation topic and depth, with failover support.
• Prompt configuration: enables capabilities through system-prompt placeholders and reusable templates.

The project's promotional image states its positioning plainly: VCP wants to give AI a world in which it can persist. The image also includes a deployment warning that deserves to be called out separately.

The warning says: be sure to read the README before deploying. VCP Agents have low-level permissions in distributed systems. Do not use any unofficial or reverse-proxy APIs (mirror-site relay APIs). With low-level monitoring permissions, an untrusted API could leak sensitive information such as interaction data, memory database contents, and keys. Non-expert users should deploy with caution. This is not boilerplate: with permissions granted to this degree, the cost of connecting to the wrong API is very real.

Many Agent frameworks on the market have different packaging, but at their core they treat AI as a callee: it wakes up when asked a question and goes to sleep after answering. The project's promotional image puts this pointedly: every conversation is a new birth, and every ending is a memory loss. VCPToolBox wants to change this default setting. It puts memory, plugins, and routing into a long-running system layer so Agents don't have to start from zero every time.

For developers, the value lies in doing less repeated work building memory and tool wiring, especially in scenarios that require multiple frontends, multiple plugins, and frequent model switching. Text-based tool calling also leaves a path for models without native function calling. For ordinary users, the concept is easy to understand, but the deployment barrier and risks are not low. If you want to use it commercially, check the license first; if you want to deploy it, read the README first, then isolate permissions, and don't connect untrusted APIs to an Agent with low-level permissions. Project URL: https://github.com/lioensky/VCPToolBox . Tweet link: https://x.com/DanKornas/status/2117362756575707628 .