| Drop a star to support Nu ⭐ | Join the Nu Discord community |
Build apps in one primitive that spans your whole stack — databases, UIs, AI agents, and services. No glue. 50x less code.
The primitive Nu is built on, in three words.
A tiny script keeps its values in memory and calls functions on them. Real apps don't stay there. Values live in a database, arrive from a browser form, render into a dashboard, get recomputed by a background job. Almost none of the code is about the values anymore — it's all interaction between systems.
Nu makes interaction the primitive.
- Ref — a name for a value, wherever it lives: a KV slot, a UI text block, an LLM endpoint, a remote object.
- Interaction — what you do with a Ref: read, write, branch, iterate, compose.
- Fabric — binds Refs to a real system. Compose as many as you need; they all speak the same primitive.
Same primitive, different system. One Ref for any resource, one Interaction for any op. Setting a browser text block is the same interaction as setting a KV slot.
Nu is the reference implementation of the interaction model in Python.
A few of Nu's fabrics — state, UIs, cluster, models. Same Python, same code shape.
Persistent state. Stores any Python type.
Reach any value by name; it survives restarts. Same code local or sharded across a cluster.
- Serverless, scales to terabytes
- Store any Python type
...
classDB(nu.Shape):
hits=nu.kv.IntRef.slot()
# +1 to a persistent counterop=DB.hits.set(DB.hits+1)
app=nu.With(
nu.kv.rocksdb_navigator(".db"),
body=nu.kv.auto_flow_atomic(op),
)
...Powered by virtuals · rdbpy · RocksDB · LMDB.
Reactive UIs from Python. No JS, no build step, no websocket you had to write.
A text block, a chart, a form — set them like variables, they render. Update the value, the browser updates itself.
- 50+ components out of the box
- Live updates for free
...
classDashboard(nu.ui.Page):
hello=nu.ui.TextRef.slot()
# renders live in the browserop=Dashboard.hello.set("Hello, browser.")
app=nu.With(
nu.ui.server(op),
)
...Same code runs local or across the cluster. No worker pool to run.
Teleport any Nu tree to any worker; it runs there and returns the result. Where it runs is a binding, not a rewrite.
- Same code, local or remote
- Distribute with a single line
...
classDB(nu.Shape):
hits=nu.kv.IntRef.slot()
# any Nu opop=DB.hits.set(DB.hits+1)
# same op — teleport it to a workerremote=nu.cluster.Teleport(op, target="gpu-0")
app=nu.With(
nu.cluster.RayCluster(),
body=remote,
)
...Powered by Ray.
One wire, N providers. Ollama, OpenAI, OpenRouter, Groq, xAI — same call.
LLM chat as a Ref. Swap the model string, keep the code. Local models and hosted APIs meet at the same interface.
- Add intelligence to any Nu app
- 7 providers out of the box
...
classBot(nu.Service):
chat=nu.llm.ChatRef.method(temperature=0.7)
# ask the modelop=Bot.chat(prompt="one-line haiku about rust")
app=nu.With(
nu.llm.ollama(Bot, host="localhost", model="qwen2.5:7b"),
body=op,
)
...Nu is batteries-included and covers the common cases. In-memory state, proxy, HTTP, Python objects, Claude Code, local parallelism — same model, same shape, same primitive as the four above.
Install, run a demo, start hacking.
Python 3.10+ · everything ships in the wheel.
pip install "nustack-py[all]"Each one boots a live browser dashboard and picks up where it left off on restart. nu demo lists them all.
| counter | sampled |
|---|---|
 A live counter, persistent across restarts. nu demo counter |  An infinite series, live-sampled into a line-chart. nu demo sampled |
| movies | |
 A movie tracker: form, filterable table, detail pages. nu demo movies |
- Read the docs — tutorials, how-tos, and the fabric reference.
- Browse examples — full source for every demo, plus more programs to steal from.
Each fabric gives your Nu app a new capability. These are the ones Nu ships with today.
| Fabric | What | Primary interaction |
|---|---|---|
| nu.kv | Persistent state. | State.movies.append(m) |
| nu.ui | Reactive web UI. | Dashboard.count.set_value(n) |
| nu.cluster | Cluster compute. | Teleport(Add(1,2), "gpu") |
| nu.llm | OpenAI-compatible chat. | Bot.chat(prompt="…") |
| nu.mem | In-memory state. | users.age.set(12) |
| nu.proxy | Fabrics over the network. | Proxy(Nav, "10.0.0.1") |
| nu.http | Nu meets the web. | Solana.get_slot() |
| nu.service | Python objects as Refs. | Calc.add(a=2, b=3) |
| nu.cc | Claude Code as a Ref. | Agent.ask(prompt="…") |
| nu.mp | Local parallel execution. | Teleport(Add(1,2), "worker") |
End-user tools written as Nu programs.
| Repo | What |
|---|---|
| nustackdev/nulog | Pure-Python, serverless logger and metrics store. Log and observe metrics from any Python code; entries persist to an embedded KV store and scale to billions, in-process. One line boots a live viewer. |
Nu is a reference implementation of the interaction model — a language-agnostic specification of what an interaction is, how Refs name locations, and how Interactions compose into programs.
nu.agents fabric — LLM authors Nu programs.
No tool-calling loop. The model's reply is a Nu tree, evaluated in the Context the agent runs in. The Refs bound in scope are the agent's surface — bind different Refs, get a different agent.
- Safe by construction. Nu's laws validate the tree before any effect fires;
With/Providescoping bounds what the model can touch. - Inspectable, replayable, diffable. A Nu program is a data structure.
Add a Nu badge to your README if you're building on Nu:
[](https://github.com/nustackdev/nu)
Considering contributing to Nu? Start by opening an issue or a PR — the codebase is small and readable, and the model is stable enough to build on.
Apache-2.0