Agent RT for TypeScript

Production AI agents without the framework tax. A lightweight, provider-neutral runtime with explicit control over tools, safety, state, and orchestration.

🌐 Website · Documentation · GitHub

Meet Artie, the running robot, and its flagship features: Guard (blocks harmful tool calls before they run), Checkpoints (resume with budgets intact), and the migration Wall (65+ repositories passing their tests after migrating).

Fully typed, with a native fetch OpenAI-compatible transport. Start with zero provider SDKs, then add tools, approvals, memory, sandboxing, retrieval, and multi-agent orchestration as needed.

Why Agent RT?

• Provider-neutral. OpenAI-compatible endpoints, OpenAI, Anthropic, or your own ModelProvider.
• Safe by construction. Tool validation, permissions, composable action blockers, approvals, deterministic PII morphing, guardrails, budgets, and deadlines are runtime primitives.
• Scales large tool catalogs. Toolbase combines local/deferred and MCP tools, safety-screens them, sends only the relevant subset per turn, and keeps tool_search available for fallback discovery.
• Small core. OpenAI, Anthropic, Agent Action Guard, and sandbox integrations are optional peers.
• Production-ready. Checkpoints, events, queues, tracing, cost ledgers, retrieval with post-retrieval reranking, and multi-agent patterns.
• Decision models. Jev-compatible models, or the OpenAI-compatible Decisions API answer typed questions that gate tool safety, tool relevance, memory, and retrieval (details).
• Easy to adopt. Migration-compatible entry points for supported subsets of OpenAI, Anthropic, OpenRouter, LangChain, Deep Agents, LlamaIndex, and OpenAI Agents code; across Agent RT's ongoing migration field testing, 65+ third-party repositories pass before and after migration.

Install

Node 20, 22, 24, or 26. Runs on Linux, Windows, macOS (Intel and Apple M-series), and Raspberry Pi.

npm install agent-rt

Provider SDKs are optional: add openai or @anthropic-ai/sdk only when you need them. For agent memory, add one of mem0ai (the default), @getzep/zep-cloud, @letta-ai/letta-client, or supermemory and call loadMemory() (details). Agent RT works with the SDK versions your project already pins; releases since mid-2024 (openai 4.48+ through 7.x, @anthropic-ai/sdk 0.23+) are regression-tested.

For per-session task lists, new TodoList().register(registry) gives a model one todo tool (details).

Migrate in one line

- import OpenAI from "openai";
+ import OpenAI from "agent-rt/openai";

Prefix a supported import with agent-rt/, for example import { ChatOpenAI } from "agent-rt/@langchain/openai". Across ongoing field testing, 65+ third-party repositories pass their own tests before and after this switch. See Migration.

Try it with no API key: npx agent-rt demo runs a scripted mock model through a prompt-injection attack, shows the built-in action blocker stopping rm -rf ~ before the tool runs, totals tokens, and prints this process's peak RAM next to a competitor's published benchmark figure.

Or let Agent RT rewrite every supported import for you (after npm install agent-rt):

npx agent-rt migrate

It runs your tests before and after the rewrite and keeps it only if nothing new fails: every test that fails afterwards must have failed before, in the same test at the same line. Otherwise every file is restored. It then prints a before-and-after card (tests, test time, peak memory, package size) and, for a GitHub repository, a pre-filled issue link to add it to the migration wall. Per-test comparison needs Vitest or Jest.

Migrated? Add the opt-in Runs on Agent RT badge to your README.

Quick start

import {
  AgentLoop,
  loadConfig,
  loadModel,
  type AgentConfig,
  type ModelMessage,
} from "agent-rt";

// ~/.agent-rt/config.yaml by default; project files require loadConfig("./agent-rt.yaml")
const provider = await loadModel(await loadConfig());

const agent: AgentConfig = {
  name: "assistant",
  instructions: "Answer clearly and concisely.",
  model: { model: "my-model" }, // a model_name from agent-rt.yaml
};

const messages: ModelMessage[] = [
  {
    role: "user",
    content: [{ type: "text", text: "What can Agent RT do?" }],
  },
];

const result = await new AgentLoop(provider).run(agent, messages);
console.log(
  result.finalResponse?.message.content
    .filter((part) => part.type === "text")
    .map((part) => part.text ?? "")
    .join("") ?? ""
);

Configure several models and vector stores in agent-rt.yaml; secrets are referenced as os.environ/NAME and nothing else is read from the environment. Bedrock Knowledge Bases uses the optional @aws-sdk/client-bedrock-agent-runtime dependency and AWS IAM credentials. See Configuration and the example file.

model_list:
  - model_name: my-model
    agent_rt_params:
      model: openai/your-model        # or anthropic/..., add api_base for compatible endpoints
      api_key: os.environ/OPENAI_API_KEY

More examples: examples/.

Learn more

• Website and documentation site
• Runtime guide — sandboxes, skills, providers, migration entry points
• Architecture and Extensions · Introduction to Agent Harnesses
• Development · Comparisons