One line to start. Any model. Tools from a type hint. Python and TypeScript, same API.
from nobelagentic import Agent, tool @tool def get_weather(city: str) -> str: """Current weather for a city.""" ... agent = Agent("You help plan trips", tools=[get_weather, "web"]) agent.run("Plan a 3-day trip to Lagos")
Most frameworks make you learn their graph, their crew, or their chain on day one. Here you start with one agent and reach for more structure only when you need it. All three levels run on the same kernel, so streaming, tracing and checkpoints behave identically everywhere.
researcher = Agent(
"Find and summarize sources",
tools=["web", "files"],
mcp=["github"],
)
researcher.run("What changed in Python 3.14?")
# v0.3 crew = Crew( agents=[researcher, writer, editor], process="sequential", memory=True, ) crew.run("Write a report on African fintech")
# v0.2 g = Graph(state=ReportState) g.node("research", researcher) g.node("write", writer) g.edge("research", "write") g.edge("write", lambda s: "research" if s.needs_more else END) g.run(ReportState(topic="..."), checkpoint="run-42")
A function with type hints and a docstring is a tool. Built-in packs load lazily, declare what credentials they need, and stay out of the context window until you ask for them.
webSearch, fetch and clean pages, raw HTTP requests.
filesRead, write, list, regex-search — sandboxed to a root.
shellCommands and in-process code. Opt-in only.
timeNow in any timezone, date math, conversions.
mathSafe expression evaluation and unit conversion.
dataJSON paths, CSV in and out, quick column stats.
Name a server and its tools appear next to your own, indistinguishable from native ones. A built-in registry covers GitHub, Postgres, Playwright, the filesystem and more; add your own with one command.
Agent(mcp=["github", "postgres"]) # or a remote server Agent(mcp=["https://tools.example.com/mcp"]) # register one for every project $ nobelagentic mcp add weather "npx -y weather-mcp"
Set an API key and the model is picked for you. Name a model and the provider is inferred from it. Swap vendors without touching your agent code.
What's here today works today. What's coming has a version on it — no vaporware, the roadmap is public.
Token deltas, tool calls and usage as one typed event stream — the same in Python and TypeScript.
Transfer the conversation to a specialist — or keep control: worker.as_tool() lets a supervisor delegate and combine.
When the model asks for three tools at once, they run at once.
Get a validated Pydantic model or zod object back, not a string to parse.
Multi-turn conversations with pluggable storage. No hand-rolled history.
Checkpoint before a risky tool runs, resume after a human approves.
Vector memory, plus retrieval over your own docs with one argument.
OpenTelemetry traces and per-run cost tracking. See exactly what your agents are doing.
Record a real run once, replay it offline in CI — assert on tool calls, outputs and cost.
The Examples page walks through each of these — complete, runnable, webhook included.
WhatsApp Cloud API webhook → agent with an order-lookup tool → reply. One file with FastAPI.
Read the recipe →Billing and tech specialists behind a triage agent that routes with handoffs and never answers itself.
Read the recipe →Long-polling bot with a live FX-rate tool. No bot framework, ~25 lines.
Read the recipe →| NobelAgentic | OpenAI Agents SDK | CrewAI | LangGraph | |
|---|---|---|---|---|
| Lines to a working agent with a tool | ~6 | ~12 | ~25 | ~40 |
| Python and TypeScript, same API | Yes | Yes | Python | Yes |
| Any provider without an adapter package | Yes | OpenAI-first | Via LiteLLM | Via LangChain |
| Built-in tool packs | 20+, growing | Few | Many | Via LangChain |
| MCP client | By name | Yes | Yes | Adapter |
| Graph, crew and single-agent in one runtime | Yes | Agent + handoffs | Crew + flows | Graph |
Line counts are from each project's own quickstart as of October 2026. Crew and Graph APIs are on the roadmap for v0.2–0.3.
An open-source (MIT) agent framework for Python and TypeScript. One kernel with three API levels — Agent, Crew, Graph — so you start with one line and add structure only when you need it.
Anthropic, OpenAI, Gemini, Ollama, Groq, OpenRouter, Together — any OpenAI-compatible host — or bring your own provider class. Set an API key and a model is picked for you; name a model and the provider is inferred.
Yes, as a client: name a registry server (mcp=["github"]), point at a remote URL, or spawn a local one. MCP server mode — your agents as tools for any MCP client — lands in v0.4.
Today you prepend recent history (the Examples page shows the pattern). Sessions with pluggable storage land in v0.2; vector memory and knowledge sources in v0.3.
Both, with the same API and the same event stream — pip install nobelagentic or npm i nobelagentic. The two packages are ports of one shared spec.
Those frameworks make you learn their abstraction on day one. Here complexity is opt-in: a working agent with a tool is ~6 lines, and crews and graphs compile to the same runtime when you grow into them — streaming, tracing and checkpoints behave identically at every level.
v0.1 is young and the API may still move before v1. What exists is tested (offline, with a scripted fake provider) and the roadmap is deliberately boring-first: structured outputs, sessions, run limits and human-in-the-loop are next.
An agent with a real tool, in the next five minutes.