We've been hacking around with LLMs for a while and have encountered a specific problem with distributed tool calling.

The Problem

When building AI agents or LLM-automations in distributed environments, you typically:

- Need to build APIs for your distributed tools

- Require load balancers in front of tool replicas (e.g. in k8s environments)

- Must refactor long-running tools to work within HTTP timeout constraints

Our Solution: AgentRPC

AgentRPC addresses these challenges by converting any function into a consumer for a distributed message queue that works via long-polling. The consumers register with a centralized server which:

- Monitors their health

- Maintains context about function schemas

## Features

The AgentRPC SDKs provide:

- A unified MCP-compatible server

- Tool definitions in an OpenAI SDK compatible format

The AgentRPC server handles:

- Load balancing

- Automatic failover

- Observability

Because tool calling happens through an async HTTP-based API, it can handle tool calls well beyond HTTP timeout limits.

We currently support TypeScript, Go, and Python natively, with more SDKs in development.

Check us out: https://agentrpc.com/

We're still early, but keen to hear any feedback!