Open Agent SDK

Minimal, production-ready TypeScript SDK for building tool-using AI agents.

Build agents with a ReAct loop, tool permissions, hooks, subagents, session persistence, and multi-provider support.

1-Minute Quickstart

npx open-agent-sdk@alpha init my-agent
cd my-agent
npm install
cp .env.example .env
npm run dev

Or with Bun:

bunx open-agent-sdk@alpha init my-agent

30-Second Demo

More runnable demos: Demo Gallery.

Why Open Agent SDK

• Production safety controls: permission modes (default, plan, acceptEdits, bypassPermissions) and per-tool gating via canUseTool.
• Agent extensibility core: hooks, skills, subagents, and MCP-compatible tool integration.
• Reproducible evaluation path: local SWE-bench and Terminal-bench harnesses in benchmark/.

See details in:

• API Reference
• SWE-bench Guide
• Terminal-bench Guide
• Benchmarks

Concepts

• Agent loop: multi-turn ReAct with tool execution.
• Tool permissions: explicit allow/deny policy hooks.
• Hooks: lifecycle/tool events for observability and control.
• Subagents: task delegation and orchestration.
• Sessions: create, save, resume, and fork conversations.

Example Gallery

• Interactive Code Agent CLI
• Quickstart Tests (basic/session/tools)
• Skill System Demo
• Structured Output Demo
• File Checkpoint Demo

Evaluation

• SWE-bench Lite smoke/batch runners: benchmark/swebench/scripts/
• Terminal-bench Harbor adapter and runbook: benchmark/terminalbench/
• Result summarization scripts and artifacts: see BENCHMARKS.md

Integrations

Current provider support in core SDK:

• OpenAI
• Google Gemini
• Anthropic

Ecosystem integrations:

• MCP server integration support
• Harbor adapter for Terminal-bench

Docs

• Homepage: https://openagentsdk.dev
• Docs: https://docs.openagentsdk.dev
• GitHub: https://github.com/OasAIStudio/open-agent-sdk
• Introduction
• Comparison with Claude Agent SDK

Monorepo Layout

packages/
  core/        # SDK implementation
  web/         # product homepage (Next.js)
  docs/        # docs site (Astro + Starlight)
examples/      # runnable examples
benchmark/     # eval harness and scripts
docs/          # engineering docs, workflows, ADRs

Development

# install dependencies
bun install

# build core package
bun run build

# run tests
bun test

# run coverage
bun test --coverage

# type check
bun run typecheck

Integration tests with real LLM APIs:

env $(cat .env | xargs) bun test

Project Status

Current release line: 0.1.0-alpha.x.

The repository is under active development. APIs may evolve before stable 1.0.0.

Contributing

Please read CONTRIBUTING.md before opening PRs.

License

MIT