JarvisMCP

The standard way to give an AI agent capabilities is one tool per endpoint. That approach collapses at scale: every tool definition is spent from the context window before the agent has done anything, so a few dozen integrations can consume thousands of tokens purely describing what is available. Worse, each new server is another process to run, another credential to distribute, and another place for a secret to leak into a model's reasoning trace.

JarvisMCP inverts the model. Instead of exposing N tools, it exposes two — one to search a ranked capability catalogue, one to execute code — and lets the agent write JavaScript against a single unified SDK. That code runs inside a workerd V8 isolate under a hard timeout, so arbitrary agent-authored programs are contained by construction. The security property that matters is the boundary: upstream credentials for every integrated system live only in trusted processes on the far side of a loopback line, and nothing the agent writes can reach across it. The isolate can ask for work to be done; it can never see the key that does it.

• 57 services and 454 discoverable methods reachable through exactly 2 tools
• Over 99% token reduction versus a conventional tool-per-endpoint server
• Arbitrary agent-written JavaScript contained in a V8 isolate under a hard timeout
• No upstream credential ever crosses into the sandbox — secrets stay in trusted processes
• Batching many calls into one execution turns per-call round-trips into a single request
• Deployed on Docker Swarm with bearer-token auth and per-key rate limiting