Durable state, zero-hop platform bindings, your framework unchanged.
Every other agent stack has a seam. Your agent runs on a compute platform, its memory lives in a database, its LLM calls go to an inference provider, its phone calls go through a telephony API. AgentSDK closes the seam.

Inference, voice, messaging, KV, and storage all through the same API key. No multi-vendor setup, no separate billing, no integration glue between services.

Your agent runs on Edge Compute in the same datacenter as inference GPUs. No cross-region hop to a cloud provider endpoint, no network latency between your logic and your model.

Voice and messaging run on Telnyx own network, not a third-party telephony API. Your agent phone calls never leave the platform.
Three primitives for durable state. Platform bindings to voice, messaging, inference, and storage. No credentials, no network calls.
Persistent message history
Every message stored in a structured table. Query by role and time. Your agent remembers context across sessions without external infrastructure.
Durable scheduled tasks
Timers that survive restarts. Schedule follow-ups, queue background work, and trust that tasks execute even after deploys.
Merge-patch state
Durable state on a single KV key. Concurrent updates merge, they dont clobber. No race conditions, no lost state.
Inference (via telnyx binding)
Call models on owned GPUs through the telnyx binding. OpenAI-compatible. Same facility as your agent code, no network hop. Accessed via this.env.TELNYX.ai.
Voice and messaging bindings
Make and take real phone calls. Send and receive SMS, MMS, and WhatsApp. Carrier infrastructure, not a third-party telephony API.
Storage bindings
KV and Cloud Storage through platform bindings. Zero-credential access from your agent code. Accessed via this.env.
One class. Three primitives. Zero-hop bindings.
import { Agent } from "@telnyx/edge-runtime";
class SupportAgent extends Agent {
async onMessage(msg) {
const res = await this.env.TELNYX.ai.openai.chat.createCompletion({
model: "glm-5.2",
messages: [...await this.messages.list(), msg],
});
await this.env.TELNYX.messages.send({
to: msg.from, from: msg.to, text: res.choices[0].message.content,
});
await this.schedule("follow-up", 86400);
}
}AgentSDK runs on Edge Compute. These are the platform primitives your agent can reach without leaving the network.
Your agents code, state, storage, model calls, phone calls, and text messages, on one platform.

Your agents code, state, storage, model calls, phone calls, and text messages, on one platform.

AgentSDK is a TypeScript base class that extends StatefulActor with three primitives: persistent message history, durable scheduled tasks, and merge-patch state. It ships inside @telnyx/edge-runtime.
LangGraph, LangChain, or any framework that runs on Node. You subclass Agent instead of StatefulActor and bring your existing agent code. The platform adds message tables, task queues, durable timers, and framework-shaped history.
Bindings are platform-level connections to Telnyx primitives. The telnyx binding gives your agent access to inference, voice, and messaging via this.env.TELNYX. KV, cloudstorage, and other bindings are accessed via this.env with no credentials to manage.
Cloudflare Agents requires their runtime and their agent model. AgentSDK runs your framework unchanged.
AgentSDK ships inside @telnyx/edge-runtime. You pay for the Edge Compute resources your agent uses.
AgentSDK is in beta. The API surface may change before GA. WebSocket support, an operating envelope, and cross-channel memory are on the roadmap.
Read the developer docs and check out the GitHub repo for examples.