Agent SDK

A composable TypeScript/Bun SDK for building LLM agents with tools, memory/context, teams, and task orchestration.

Runtime Support

• Node.js 18.18+
• Package outputs: ESM + CommonJS + TypeScript declarations

Install

npm install agent-sdk

For local development in this repository:

bun install

Build

npm run build

Native Code Understanding

Code understanding now runs as a native in-process capability when enabled.

• jcodemunch and native-dart providers both resolve to NativeCodeUnderstandingAdapter.
• native-dart applies a Dart language preference for ranking, while keeping the same core capability surface.
• Stdio transport settings do not switch runtime behavior to MCP for these providers.

Operational limits are enforced to keep indexing bounded and predictable:

• maxFiles (default 2500)
• maxFileSizeBytes (default 524288)
• maxTotalBytes (default 67108864)
• source roots are constrained to authorized directories

These limits can be tuned through code-understanding configuration.

Import

import { createOrchestrator } from 'agent-sdk'

CommonJS:

const { createOrchestrator } = require('agent-sdk')

CLI

agent run "Summarize this repository"
agent run --file prompts.txt --output json
agent config init

Examples

bun run examples:smoke
bun run lab:lovable
bun run lab:web-ui

lab:lovable defaults to NVIDIA's OpenAI-compatible endpoint (https://integrate.api.nvidia.com/v1) and model z-ai/glm4.7. It now includes a concrete AI app-builder scenario that writes outputs inside a per-run sandbox under examples-output/lovable-lab/, with app files in apps/todo-app-<run-id>/ and run artifacts in artifacts/.

Set credentials before running:

set NVIDIA_API_KEY=your_key_here

lab:web-ui starts a local Lovable-style demo UI at http://localhost:8788:

• Left pane: prompt/chat input for app-building requests.
• Right pane: live execution trace (tool calls/results, assistant deltas) plus preview iframe.
• Generated app files are sandboxed per session under examples-output/lovable-web-studio/apps/.

Docs

bun run docs:api

Benchmarks

npm run bench:resilience
npm run bench:tasks-dag-stress
npm run bench:code-understanding-large-repo

Project Overview

Read the full architecture and usage guide here:

• docs/project-overview.md
• docs/frontend-integration-guide.md