Install

Install the Nanocodex CLI on macOS or Linux:

curl -fsSL https://nanocodex.paradigm.xyz | bash

Or add the Rust SDK to an application:

cargo add nanocodex

Switch the installed CLI between builds:

nanocodex update                 # latest stable release
nanocodex update 0.2.0           # exact release, including downgrades
nanocodex update --nightly       # latest nightly
nanocodex update --pr 50         # verified on-demand PR artifact
nanocodex update --path ./nanocodex  # trusted local binary

Downloaded builds are retained under ~/.nanocodex/versions. Running
nanocodex update 0.2.0 again switches to the cached binary without another
download. A stable launcher keeps nanocodex update available even while an
older binary is active, and ~/.nanocodex/current points to the selected
version.

PR artifacts require an authenticated gh CLI and an already completed
on-demand artifact workflow for that PR.

Minimal API Example

use nanocodex::{Nanocodex, OpenAi};

let openai = OpenAi::new(std::env::var("OPENAI_API_KEY")?)?;
let (agent, mut events) = Nanocodex::builder(openai)
    .instructions(
        "You are a Rust coding agent. Make focused changes, preserve unrelated work, \
         and run relevant tests before finishing.",
    )
    .workspace(std::env::current_dir()?)
    .build()?;

let event_task = tokio::spawn(async move {
    while let Some(event) = events.recv().await {
        eprintln!("event {}: {:?}", event.seq, event.kind);
        if event.kind.is_terminal() {
            break;
        }
    }
});

// Alternative: stream this turn's response as it arrives:
// use futures_util::StreamExt;
// use nanocodex::agent::events::{AgentEventData, AssistantEvent};
// let mut turn = agent.prompt("Find and fix the failing parser test.").await?;
// while let Some(event) = turn.next().await {
//     if let AgentEventData::Assistant(AssistantEvent::Delta(delta)) = event.data()? {
//         print!("{}", delta.text);
//     }
// }
let result = agent
    .prompt("Find and fix the failing parser test.")
    .await?
    .await?;

event_task.await?;
println!("{}", result.final_message());

The first await accepts and orders the prompt. The second waits for its typed
TurnResult. Follow-on prompts automatically reuse the agent's retained
history, WebSocket, tools, shell sessions, and prompt-cache identity.
agent.clone() is a cheap handle to that same session; the independently
returned AgentEvents stream is the session-wide event firehose.

Nanocodex supports gpt-5.6-sol (the default) and gpt-5.6-luna. Select the
model with .model(Model::Luna) when creating an agent. The model is fixed for
that thread: switching later would invalidate the provider checkpoint and
require an inefficient replay of the complete retained context.

Voice: devices or Unix pipes

The non-TUI desktop example owns the default microphone and speaker directly
in Rust, using the same VoiceSessionBuilder as the production TUI:

nanocodex auth login # once; shares ~/.codex/auth.json with Codex
cargo run -p nanocodex-examples --bin voice

The lower adapter leaves device and media ownership outside Nanocodex. It reads
24 kHz mono PCM16 little-endian audio from stdin, writes the same format to
stdout, and keeps transcripts and agent events on stderr:

cargo run -p nanocodex-examples --bin realtime-pipe < microphone.pcm > speaker.pcm

# Equivalently, compose any live capture/decoder and playback/encoder:
capture-s16le | cargo run --quiet -p nanocodex-examples --bin realtime-pipe | play-s16le

Both retain one coding-agent session. A spoken request starts work while idle;
a follow-up received during that work atomically steers the active turn at its
next safe model boundary. Both use shared Codex/ChatGPT subscription auth, not
an API key. Set NANOCODEX_AUTH_FILE to override the normal Codex credential
path.

Thesis

Small, excellent building blocks

Agent infrastructure is easier to understand and reuse when each piece has a
sharp owner and a useful API of its own. An OpenAI client should work without
an agent loop. Tools should work without a CLI. The high-level agent should
compose those pieces rather than hide another implementation of them.

Nanocodex makes a small number of deliberate choices—Rust, Tower, typed
protocols, owned lifecycle state, and builder APIs—then keeps the boundaries
boring.

The model and harness are co-designed

We do not try to outsmart behavior that frontier models and Codex already make
explicit. Context management, AGENTS.md, compaction, cache identity, tool
shapes, continuation, reconnect replay, cancellation, and process cleanup are
parts of the model-facing contract.

Nanocodex carries those invariants into a smaller, library-first API while
leaving application policy with the caller.

Evidence over intuition

Representative cargo bench workloads, OpenTelemetry traces, differential
tests, and end-to-end evals keep the harness honest. The goal is simple: normal
agent turns should be model- and network-latency bound, with token usage and
estimated USD cost visible at the same typed boundary as the result.

Components

nanocodex                         Alloy-style facade and prelude
├── agent                         nanocodex-agent
│   ├── oai                       nanocodex-oai-api
│   └── tools                     nanocodex-tools
│       └── macros                nanocodex-tools-macros
├── oai                           nanocodex-oai-api
├── tools                         nanocodex-tools
└── observability                 nanocodex-observability (optional)

The facade provides the canonical common imports. Each lower crate is also
designed to be useful directly, without importing the higher orchestration
layer.

nanocodex

The thin facade reexports the golden agent path at the crate root and keeps
detailed APIs under nanocodex::agent, nanocodex::oai, and
nanocodex::tools. Its prelude contains only the common types needed to build
an agent.

Facade guide ·
API documentation

nanocodex-agent

The batteries-included lifecycle: an owned private driver, a cheap cloneable
Nanocodex handle, typed Turn and TurnResult values, and an optional event
stream. It owns prompt ordering, the tool loop, AGENTS.md discovery,
compaction timing, cancellation, snapshots, and branching through spawn,
fork, and fork_from.

Callers never pass previous messages, response IDs, or tool results back into
the agent.

Agent guide ·
API documentation

nanocodex-oai-api

The complete OpenAI boundary: API-key and ChatGPT authentication, typed
Responses protocol values, a persistent WebSocket transport, client-owned
context, continuation and replay, automatic pricing, and a generic Tower
client.

Its standalone OpenAi -> Session -> ResponseTurn -> Response path provides a
managed conversation without taking on agent policy. Custom Tower layers and
services remain concrete and nameable—no boxing or global client is required.

OpenAI API guide ·
API documentation

nanocodex-tools

The model-facing tool runtime: the Tool contract, heterogeneous Tools
registry, standard workspace tools, shell and process lifecycle, Code Mode,
deferred tool_search, remote dispatch, and MCP. MCP is always available on
native targets.

Applications can implement Tool directly or use the reexported #[tool]
macro. The separate nanocodex-tools-macros package exists only for Rust's
procedural-macro boundary.

Tools guide ·
API documentation

nanocodex-observability

Application-owned tracing and OpenTelemetry setup for the data already flowing
through the agent. It provides structured lifecycle, model, tool, usage, cost,
cache, and latency telemetry without changing the core runtime path.

Enable the facade's observability feature or depend on the component
directly.

Observability guide ·
API documentation

Experimental components

Components whose public contracts are still maturing live under
crates/experimental/:

The CLI is a consumer of these crates. Voice and VM-backed tools remain thin,
opt-in adapters over the stable library contracts.

CLI and language bindings

The CLI/TUI, Python package, Node/browser package, React bindings, and examples
are thin consumers of the same owned session API. They do not define a second
agent protocol.

Examples · JavaScript ·
Python · Web

VM-backed tools

Normal TUI and one-shot sessions keep host workspace tools by default. They can
instead route exec_command, write_stdin, apply_patch, and view_image
through one retained VM:

just build-vm-guest
nanocodex \
  --vm .nanocodex/vm/session-rootfs.ext4 \
  --vm-guest-runtime target/aarch64-unknown-linux-musl/debug/nanocodex-vm-guest \
  --vm-workspace /app
nanocodex run "inspect the repository" \
  --vm .nanocodex/vm/session-rootfs.ext4 \
  --vm-guest-runtime target/aarch64-unknown-linux-musl/debug/nanocodex-vm-guest \
  --vm-workspace /app

Directory roots may instead contain
/usr/local/bin/nanocodex-vm-guest. See the
VM guide for image preparation, lifecycle, egress, Linux
requirements, and macOS signing.

Documentation

• Facade API
• Migration from 0.2.x
• Examples
• Benchmarks and retained measurements
• VM-backed tools and egress

License

Licensed under either of:

• Apache License, Version 2.0
• MIT License

at your option.