AIPex/packages/core

@aipexstudio/aipex-core

Platform-agnostic TypeScript building blocks for creating streaming, tool-using AI agents.

@aipexstudio/aipex-core wraps @openai/agents and adds a few opinionated layers that are useful for productizing agents:

• A stable event stream (AgentEvent) that UIs can consume
• Optional conversation/session management (persistence, listing, forking)
• Optional conversation compression for long-running sessions
• A context system for attaching external data to a turn (ContextManager + ContextProvider)
• Tool utilities, including a schema-first ToolRegistry

Why a separate "core" package?

• Platform boundary: no Chrome APIs, no Node-only assumptions. This keeps core usable in browsers, servers, and extensions.
• Streaming-first: instead of returning a single string, the agent emits structured events (text deltas, tool lifecycle, metrics, errors).
• Composable integrations: persistence, context providers, and UI are all optional and pluggable.

Core concepts

Agent + event stream

AIPex.chat() returns an AsyncGenerator<AgentEvent>. Typical consumers:

• CLI: print content_delta as it streams
• UI: render streaming text, show tool calls, attach contexts, show metrics

Key event types include:

• content_delta: incremental text tokens
• tool_call_start / tool_call_complete / tool_call_error
• tool_call_args_streaming_start / tool_call_args_streaming_complete (best-effort, provider-dependent)
• session_created / session_resumed (when conversation management is enabled)
• metrics_update / execution_complete
• contexts_attached / context_error

Tools

There are two complementary ways to work with tools:

1. Direct function tools via tool() (re-exported from @openai/agents)

• Define name/description + a Zod schema
• Provide an execute() implementation

2. Registry-managed tools via ToolRegistry

• Register a UnifiedToolDefinition (schema + handler(input, ctx))
• Execute locally (registry.execute(...))
• Convert to OpenAI-compatible function tools (registry.toOpenAIFunctions())

Built-in tools shipped by core:

• calculatorTool: basic arithmetic
• httpFetchTool: simple HTTP GET fetch (returns JSON when content-type is JSON)

Conversations, sessions, and compression

When conversation management is enabled (default), the agent creates a session and persists turn items via a SessionStorageAdapter.

• Disable sessions by setting conversation: false
• Persist sessions by providing your own storage adapter
• Fork sessions (branch conversation history) with ConversationManager.forkSession(...)
• Compress sessions using ConversationCompressor to summarize older turns

Contexts

A context is an additional piece of information you attach to a single turn (e.g. current page text, a selected tab, a file, a screenshot, etc.).

• Implement a ContextProvider for your environment
• Register providers with a ContextManager
• Attach contexts to a message via ChatOptions.contexts (either full Context objects or context IDs)

Plugins

Plugins are a lightweight way to observe and customize runtime behavior without forking the agent. They can:

• mutate input before a turn (beforeChat)
• observe tool lifecycle events (onToolEvent)
• observe per-turn usage metrics (onMetrics)
• observe the final response (afterResponse)

Installation

npm install @aipexstudio/aipex-core
# or
pnpm add @aipexstudio/aipex-core

Optional peer dependencies (model providers)

@aipexstudio/aipex-core integrates well with the AI SDK ecosystem. Install the provider(s) you plan to use:

• @ai-sdk/openai
• @ai-sdk/google
• @ai-sdk/anthropic
• @openrouter/ai-sdk-provider

Usage

1) Basic streaming chat

import { google } from "@ai-sdk/google";
import { AIPex, aisdk } from "@aipexstudio/aipex-core";

const agent = AIPex.create({
  instructions: "You are a helpful assistant.",
  model: aisdk(google("gemini-2.5-flash")),
});

for await (const event of agent.chat("Say hello in one sentence.")) {
  if (event.type === "content_delta") process.stdout.write(event.delta);
}

2) Using built-in tools (calculatorTool, httpFetchTool)

import { google } from "@ai-sdk/google";
import { AIPex, aisdk, calculatorTool, httpFetchTool } from "@aipexstudio/aipex-core";

const agent = AIPex.create({
  instructions: "Use tools when appropriate.",
  model: aisdk(google("gemini-2.5-flash")),
  tools: [calculatorTool, httpFetchTool],
});

for await (const event of agent.chat("What is 15 * 234?")) {
  if (event.type === "content_delta") process.stdout.write(event.delta);
  if (event.type === "tool_call_start") {
    console.log("\n[tool_call_start]", event.toolName, event.params);
  }
  if (event.type === "tool_call_complete") {
    console.log("\n[tool_call_complete]", event.toolName, event.result);
  }
}

3) Define a custom tool (schema-first with Zod)

import { z } from "zod";
import { tool } from "@aipexstudio/aipex-core";

export const weatherTool = tool({
  name: "get_weather",
  description: "Get the weather for a city (demo tool).",
  parameters: z.object({
    city: z.string().describe("The city name"),
  }),
  execute: async ({ city }) => {
    return `The weather in ${city} is sunny and 72°F`;
  },
});

4) Sessions: continue a conversation with sessionId

By default, each chat() call starts a new session. To continue a conversation, pass the sessionId returned in the session_created event:

import { google } from "@ai-sdk/google";
import { AIPex, aisdk } from "@aipexstudio/aipex-core";

const agent = AIPex.create({
  instructions: "You remember things within a session.",
  model: aisdk(google("gemini-2.5-flash")),
});

let sessionId: string | undefined;

for await (const event of agent.chat("My name is Alice.")) {
  if (event.type === "session_created") sessionId = event.sessionId;
}

if (sessionId) {
  for await (const event of agent.chat("What is my name?", { sessionId })) {
    if (event.type === "content_delta") process.stdout.write(event.delta);
  }
}

5) Conversation compression (summarize older turns)

import { google } from "@ai-sdk/google";
import { AIPex, aisdk } from "@aipexstudio/aipex-core";

const model = aisdk(google("gemini-2.5-flash"));

const agent = AIPex.create({
  instructions: "You are a helpful assistant.",
  model,
  compression: {
    model,
    summarizeAfterItems: 20,
    keepRecentItems: 10,
    maxSummaryLength: 500,
  },
});

6) Attaching contexts to a turn

import { google } from "@ai-sdk/google";
import { AIPex, aisdk, ContextManager, type Context } from "@aipexstudio/aipex-core";

const contextManager = new ContextManager();

const agent = AIPex.create({
  instructions: "Use attached contexts when answering.",
  model: aisdk(google("gemini-2.5-flash")),
  contextManager,
});

const pageContext: Context = {
  id: "page:example",
  type: "page",
  providerId: "manual",
  label: "Example page",
  value: "This page describes the AIPex Core API.",
  timestamp: Date.now(),
};

for await (const event of agent.chat("Summarize the page.", { contexts: [pageContext] })) {
  if (event.type === "content_delta") process.stdout.write(event.delta);
}

7) Plugins

import { google } from "@ai-sdk/google";
import { AIPex, aisdk, type AgentPlugin } from "@aipexstudio/aipex-core";

const loggerPlugin: AgentPlugin = {
  id: "logger",
  hooks: {
    beforeChat: (payload) => {
      console.log("[beforeChat]", payload.input);
      return payload;
    },
    onToolEvent: ({ event }) => {
      if (event.type === "tool_call_start") {
        console.log("[tool]", event.toolName, event.params);
      }
    },
  },
};

const agent = AIPex.create({
  instructions: "You are a helpful assistant.",
  model: aisdk(google("gemini-2.5-flash")),
  plugins: [loggerPlugin],
});

8) ToolRegistry (register and execute tools dynamically)

import { z } from "zod";
import { ToolRegistry } from "@aipexstudio/aipex-core";

const registry = new ToolRegistry();

registry.register({
  name: "echo_tool",
  description: "Echo text back",
  schema: z.object({ text: z.string() }),
  handler: ({ text }) => `echo:${text}`,
});

const result = await registry.execute("echo_tool", { text: "hello" }, { sessionId: "s1" });
console.log(result); // "echo:hello"

API reference

AIPex

• AIPex.create(options: AIPexOptions): AIPex
• agent.chat(input: string, options?: ChatOptions): AsyncGenerator<AgentEvent>
• agent.getConversationManager(): ConversationManager | undefined
• agent.getContextManager(): ContextManager | undefined

ChatOptions

• sessionId?: string - resume an existing session
• contexts?: Context[] | string[] - attach contexts for this turn (IDs require contextManager)

AgentEvent (stream contract)

export type AgentEvent =
  | { type: "session_created"; sessionId: string }
  | { type: "session_resumed"; sessionId: string; itemCount: number }
  | { type: "content_delta"; delta: string }
  | { type: "tool_call_args_streaming_start"; toolName: string }
  | { type: "tool_call_args_streaming_complete"; toolName: string; params: unknown }
  | { type: "tool_call_start"; toolName: string; params: unknown }
  | { type: "tool_call_complete"; toolName: string; result: unknown }
  | { type: "tool_call_error"; toolName: string; error: Error }
  | { type: "contexts_attached"; contexts: Context[] }
  | { type: "contexts_loaded"; providerId: string; count: number }
  | { type: "context_error"; providerId: string; error: Error }
  | { type: "metrics_update"; metrics: AgentMetrics }
  | { type: "error"; error: AgentError }
  | { type: "execution_complete"; finalOutput: string; metrics: AgentMetrics };

AIPexOptions

Key fields:

• instructions: string
• model: AiSdkModel
• tools?: FunctionTool[]
• maxTurns?: number
• conversation?: false (disable session management)
• storage?: SessionStorageAdapter
• compression?: { model: AiSdkModel; summarizeAfterItems?; keepRecentItems?; maxSummaryLength? }
• conversationManager?: ConversationManager (advanced override)
• contextManager?: ContextManager
• plugins?: AgentPlugin[]

Conversations

• ConversationManager: createSession, getSession, saveSession, deleteSession, listSessions, forkSession, compressSession, getSessionTree
• Session: implements the OpenAI Agents Session interface and adds helpers like getItemCount, getSummary, fork, addMetrics, getSessionMetrics
• SessionStorage: adapts a KeyValueStorage<SerializedSession> to a SessionStorageAdapter
• ConversationCompressor: summarizes older turns and keeps the most recent items

Contexts

• ContextManager: register/unregister providers, list/query contexts, fetch by ID, watch providers
• ContextProvider: interface for pluggable context sources
• Context: { id, type, providerId, label, value, metadata?, timestamp? }

Tools

• tool(...): re-export from @openai/agents to define FunctionTools with a Zod parameters schema
• ToolRegistry: registry for schema-first tools + execution + conversion to function tools
• Built-ins: calculatorTool, httpFetchTool

Storage

• KeyValueStorage<T>: save, load, delete, listAll, query, watch
• InMemoryStorage<T>: reference implementation of KeyValueStorage

Development (monorepo)

From the repository root:

pnpm --filter @aipexstudio/aipex-core build
pnpm --filter @aipexstudio/aipex-core typecheck
pnpm --filter @aipexstudio/aipex-core test

Run the included examples:

pnpm --filter @aipexstudio/aipex-core example:basic
pnpm --filter @aipexstudio/aipex-core example:fork
pnpm --filter @aipexstudio/aipex-core example:metrics

License

MIT