@ridit/ai

@ridit/ai

your agents, in minutes.

build AI agents that remember things, use tools, persist sessions, and work in teams. works with any model. no magic, no black boxes. just agents.

install

npm install @ridit/ai
# or
bun add @ridit/ai

what it does

• any model — anthropic, openai, groq, google, ollama, openrouter
• memory — inject context into any session
• sessions — persist to disk or localStorage.
• compaction — context too long? it summarizes itself. automatically.
• tools — files, bash, memory, sub-agents, and more

quick start

import { buildProvider, runLLM } from "@ridit/ai";
import { FileWriteTool, ThinkTool } from "@ridit/ai/tools";

const provider = buildProvider({
  provider: "anthropic",
  model: "claude-sonnet-4-20250514",
  apiKey: process.env.ANTHROPIC_API_KEY,
});

const { text, session } = await runLLM({
  prompt: "create a hello world python script",
  provider,
  tools: { ThinkTool, FileWriteTool },
});

console.log(text);

that's it. agent runs, uses tools, returns text and the session.

providers

import { buildProvider } from "@ridit/ai";

buildProvider({
  provider: "anthropic",
  model: "claude-sonnet-4-20250514",
  apiKey: "...",
});
buildProvider({ provider: "openai", model: "gpt-4o", apiKey: "..." });
buildProvider({
  provider: "groq",
  model: "llama-3.3-70b-versatile",
  apiKey: "...",
});
buildProvider({ provider: "google", model: "gemini-2.0-flash", apiKey: "..." });
buildProvider({ provider: "ollama", model: "llama3.2" }); // no key needed
buildProvider({
  provider: "openrouter",
  model: "meta-llama/llama-3.3-70b-instruct",
  apiKey: "...",
});

sessions

sessions are opt-in. no storage passed = runs in memory, nothing saved. your call.

node

import { createFileStore } from "@ridit/ai/utils";

const store = createFileStore({ sessionsDir: "./sessions" });

const { session } = await runLLM({ prompt: "hey", provider, store });

// resume later
const { text } = await runLLM({
  prompt: "what did i say before?",
  provider,
  session,
  storage,
});

browser

const storage = {
  async save(session) {
    localStorage.setItem(session.id, JSON.stringify(session));
  },
  async load(id) {
    const s = localStorage.getItem(id);
    return s ? JSON.parse(s) : null;
  },
  async list() {
    return [];
  },
};

const { text, session } = await runLLM({ prompt: "hi", provider, storage });

bring your own adapter. redis, supabase, sqlite — whatever you want.

tools

import {
  BashTool, // run shell commands
  FileReadTool, // read files
  FileWriteTool, // write files
  FileEditTool, // patch files
  GlobTool, // find files by pattern
  GrepTool, // search inside files
  ThinkTool, // internal reasoning step
  MemoryReadTool,
  MemoryWriteTool,
  MemoryEditTool,
  RecallTool, // semantic memory recall
} from "@ridit/ai/tools";

Create yours too!

compaction

when sessions get long, @ridit/ai summarizes the history and compacts it automatically before the next call. you don't have to think about it.

system prompts

const { text } = await runLLM({
  prompt: "review this PR",
  provider,
  system: "you are a senior typescript engineer. be direct. no fluff.",
});

api

buildProvider(config)

field	type	required
provider	"anthropic" | "openai" | "groq" | "google" | "ollama" | "openrouter"	✅
model	string	✅
apiKey	string	for hosted providers
baseURL	string	for ollama / custom endpoints

runLLM(options)

field	type	description
prompt	string	user message
provider	LanguageModel	from buildProvider()
system	string	system prompt
tools	object	tool map
session	Session	resume a session
storage	SessionStorage	persistence adapter
memoryContent	string	memory to inject
steps	number	max agentic steps (default: 100)
onToolCall	function	intercept before tool runs
onToolResult	function	observe tool output
abortSignal	AbortSignal	cancel in-flight requests

built with

Vercel AI SDK — model routing, tool calling, streaming

history

@ridit/ai started as the core of Milo — a terminal AI agent/pet. after building out memory, sessions, compaction, and multi-agent support there, it made sense to pull it out into a proper framework anyone could use.

if you want to see what you can build with it, go look at Milo.

license

MIT © Ridit Jangra

made with 💕