AgentRPC

AgentRPC: Let Your AI Agents Call Functions Anywhere

AI agents need to call functions. Sometimes those functions live behind a firewall, inside a Kubernetes cluster, or on a private VPC that the agent can't reach. AgentRPC solves this by providing a hosted RPC layer that bridges network boundaries. Your agent calls a function through AgentRPC's server; AgentRPC routes it to whatever machine registered that function. No open ports. No VPN configuration. No firewall rules.

The Core Problem It Solves

Most agent frameworks assume your tools run on the same machine or are accessible via public APIs. Reality is different. Your database sits in a private VPC. Your internal APIs require VPN access. Your ML models run on a GPU server with no public endpoint. AgentRPC lets you register functions from any of these locations and expose them to your agents through a single hosted server.

The long-polling approach matters for AI workflows. Standard HTTP requests time out after 30-60 seconds. An agent task might take 5 minutes: query a database, process results, generate a report. AgentRPC's SDKs use long polling to keep connections alive beyond HTTP timeout limits. Your function runs as long as it needs, and the agent gets the result when it's ready.

How It Works in Practice

Setup takes two steps. First, register your function with AgentRPC from whatever machine hosts it:

typescript
import { AgentRPC } from '@agentrpc/sdk';
const rpc = new AgentRPC({ apiSecret: 'your-key' });
rpc.register({ name: 'queryDatabase', handler: async (params) => { / your logic / } });
await rpc.listen();

Second, your agent calls that function through AgentRPC's API or MCP server. The SDK handles routing, health monitoring, and retries. Your function never needs a public endpoint.

Multi-Language Support

SDKs ship for TypeScript, Go, and Python, with .NET in development. You register functions in whichever language your backend uses. A Python data processing pipeline, a Go microservice, and a TypeScript API handler all become available to your agents through the same RPC layer.

This matters for teams with polyglot architectures. Your ML team writes Python. Your backend team writes Go. Your frontend team writes TypeScript. AgentRPC lets AI agents call functions in all three without any team changing their stack.

MCP and OpenAI Compatibility

AgentRPC works with MCP (Model Context Protocol) and OpenAI SDK-compatible agents out of the box. The TypeScript SDK includes a built-in MCP server that exposes registered tools to any MCP-compatible host, including Claude Desktop and Cursor. OpenAI-compatible tool definitions also work with Anthropic, LiteLLM, and OpenRouter.

This makes AgentRPC a complement to MCP, not a competitor. MCP defines how agents discover and call tools. AgentRPC handles the network plumbing to make those calls work across private networks.

Observability

The hosted platform provides tracing, metrics, and event logging for every function call. You can see which functions are being called, how long they take, which ones fail, and why. Health tracking monitors registered functions and automatically fails over to healthy instances if one goes down.

Open Source

AgentRPC's SDKs and core components are open-source under the Apache 2.0 license, hosted on GitHub. The hosted RPC server at api.agentrpc.com is a managed service that handles function registration, health monitoring, and routing.

Limitations

AgentRPC is a relatively new project with a small user community. Public discussion of production deployments is limited, and real-world case studies are scarce. The documentation covers basics well but lacks depth on advanced deployment patterns, security hardening, and scaling.

For simpler scenarios where your functions are publicly accessible, AgentRPC adds unnecessary complexity. Direct HTTP calls or standard function calling works fine. AgentRPC earns its keep only when network boundaries are the actual problem.

The managed server adds a network hop. For functions that take milliseconds, this latency overhead matters. For agent tasks that take seconds or minutes, it's negligible.

Step 1: Sign Up and Install

Create an account at agentrpc.com and install the AgentRPC SDK using npm or pip. The installation includes CLI tools and development libraries.

Step 2: Initialize Your Project

Run 'agentrpc init' to create a new project with sample agent configurations. Choose from templates for common use cases like customer service or workflow automation.

Step 3: Define Your Agents

Create agent definitions using the AgentRPC schema. Define capabilities, communication protocols, and dependencies between agents.

Step 4: Set Up Communication Channels

Configure RPC channels between your agents using the built-in networking layer. AgentRPC handles serialization, routing, and error handling automatically.

Step 5: Deploy and Monitor

Deploy your agent network using the AgentRPC runtime. Use the web dashboard to monitor agent communication and performance in real-time.

Step 6: Scale Your System

Add new agents or scale existing ones using the auto-scaling features. AgentRPC automatically handles load balancing and resource allocation.