Neutron is a self-hosted agent harness for orchestrating long-lived Claude Code sessions. It adds what a raw CLI session lacks on its own: persistent memory, projects, scheduled and autonomous jobs, reminders, and an always-on web interface (plus a mobile app). You run it yourself, backed by your own Claude subscription.
curl -fsSL https://neutronagent.ai/install.sh | sh Pre-release software, under active development. Star the repo →
Claude Code does the thinking; Neutron is everything around the thinking. It spawns and supervises your Claude Code sessions, routes each turn to the right one, and wraps them in durable state, so your agent remembers you, works on a schedule, runs jobs while you sleep, and is reachable from anywhere.
Its job is orchestration, not inference: persistent memory, projects, scheduled and autonomous work, reminders, and the surfaces you reach it through. You run it yourself, on your own machine, backed by your own Claude subscription.
The primary interface is the bundled web or mobile app, with Telegram as an optional add-on.
One warm, long-lived Claude Code session per project, context lives in the process, so it never re-feeds history. Sessions persist for days. A watchdog respawns any that wedge or crash, with full context preserved. Talk to any of them from your phone.
You bring your own Claude subscription. The claude binary owns the model relationship, Neutron never holds your credentials. And the substrate is swappable: point it at another model's API, or a local open-weight model for true airgap.
Every turn quietly extracts entities, facts, and relations into a local knowledge graph, so the agent knows you from session one and sharpens over time. It's plain markdown plus a graph on your own disk, inspectable, and behind a privacy gate. Persistent memory is powered by the excellent GBrain project by Garry Tan.
Cores are business logic that runs on top, eight ship free, and the ecosystem is open. Neutron Connect lets you invite other people into your projects, with chat and agent context shared across independently run instances.
One command bootstraps everything: the runtime, a boot/crash service, the neutron CLI, scheduled backups. It auto-starts and hands you a chat URL at localhost:7800.
Runs on macOS or Linux. Your data lives in a single folder you own and can back up by copying.
A short in-chat interview asks who you are and what you're working on, optionally imports your ChatGPT / Claude history, and generates your preferred agent personality.
Paste a token from claude setup-token and Neutron runs on your own Claude subscription. (An API key works too.)
A fleet of agents handles your projects: drafting email, running research, shipping code, firing context-aware reminders, autonomously, on schedules you set once.
It pauses and asks when a decision deserves your eyes. Everything else, it just does.
A Core is Neutron's plugin unit, a package that declares its tools, capabilities, and data namespace, capability-gated on every call. Eight ship free and bundled. The ecosystem is open: install any Core you can bun install, including your own and paid third-party Cores.
Notes, tasks, reminders, calendar (Google), email (Gmail), research, code-gen (autonomous build → review → merge), and in-chat agent settings. A working personal agent system the day you install.
Write your own, pull the community's, or buy a paid third-party Core. You install it the same way you'd install any package.
Each Core gets its own data namespace and a declared set of capabilities. The runtime enforces the gate on every tool call, so a Core only ever touches what it asked for and you approved.
Neutron Connect is like Slack Connect for AI agents. A shared project that spans independently run Neutron instances, with chat and agent context shared across the boundary. One owner, any number of collaborators, each with their own memory scope. Any instance can self-host a Connect node, free, forever.
Collaborators run their own Neutron. A shared project lets their turns route into the owner's project session, with server-side identity resolution. Everyone keeps their own install, data, and credentials.
The only capability axis is memory scope. Activity replicates through a per-project event log with a default-deny allowlist, so people only see what they've been granted, and your private memory stays yours.
A Connect node mounts the cross-instance API and a public accept page. Run it yourself alongside your gateway. No central server, no middleman, free forever.
curl -fsSL https://neutronagent.ai/install.sh | sh neutron CLI, and scheduled data backups, then auto-starts the server.
localhost:7800, asks who you are, optionally imports your AI history, and connects your Claude subscription.
~/neutron folder holds code and data. Projects are plain git folders; memory is plain markdown plus a local graph. Backup is copying the folder.
Prerequisites: Bun (the installer bootstraps it), the claude CLI authenticated with claude setup-token (or an Anthropic API key), and an optional Telegram bot from @BotFather. The gbrain memory engine is installed automatically.
Pre-release. Neutron is under active development and not ready to rely on yet, things are incomplete and will change. Star the repo to be notified when the first stable release lands.
Everything lives in plain files and a SQLite database under your home directory. Projects are git folders; memory is markdown plus a local graph. Nothing is sent anywhere except the model calls you authorize. Neutron never trains on your data.
Neutron runs your turns through the claude CLI under your own subscription. The binary holds every credential; Neutron only feeds a turn in and takes a reply out. Swap in another model, or a local open-weight one, whenever you want.
Apache 2.0. Read the code, run it yourself, audit exactly what it does, change anything. No telemetry you didn't opt into, no license checks, no phone-home. The whole runtime ships in the repo.
Point the substrate at a local open-weight model and nothing leaves your machine at all. The gateway binds loopback by default, it's off your network out of the box.
Founded by Ryan Junee. Three exits (Omnisio to YouTube, Parsable to CAI Software, Ostro Health to Veeva). Engineering deep-dives on the substrate, the build loop, and the memory layer at ryanjunee.com/posts.
Claude Code is the substrate Neutron runs on, Neutron spawns and supervises real claude sessions. The difference is everything around them: Neutron runs a whole fleet (one persistent session per project), gives them long-term memory, lets them talk to each other and run on schedules, and makes the whole thing reachable from your phone over Telegram, web, and a mobile app. It's the harness that turns a single coding agent into something that runs your coding, business, and life. Cursor is an editor; ChatGPT is a chat tab; Neutron is infrastructure you host.
Yes, Neutron is self-hosted, and that's the point: your data stays on your disk, on your machine. One command installs it (it bootstraps the runtime, a boot/crash service, the CLI, and backups, then auto-starts). You connect your own Claude subscription with claude setup-token. The gateway binds loopback by default, so it's off your network out of the box.
A Core is Neutron's plugin unit, a package that declares its tools, capabilities, and data namespace, capability-gated on every call. Eight ship free and bundled (notes, tasks, reminders, calendar, email, research, code-gen, agent-settings). The ecosystem is open: install any Core you can bun install, including your own and paid third-party Cores. The runtime cannot tell a free Core from a paid one, no license checks, ever.
You self-host, so there's no third party in the loop. Your data lives in plain files and a SQLite database under your home directory. Nothing is sent anywhere except the model calls you authorize, and the gateway binds loopback only. Memory writes pass a privacy gate that refuses any content whose origin isn't you or a collaborator you allowlisted. For the strictest setups, point the substrate at a local open-weight model and nothing leaves your machine at all.
Yes. Claude Code is the default substrate, not a lock-in. The runtime talks to a formal substrate interface, and backends are adapters, an OpenAI backend and a Codex-CLI backend already conform. Point Neutron at another vendor's API, or at your own GPU box serving open weights via vLLM, and it's just a new adapter. A local open-weight model means true airgap.
Neutron is free and open source under Apache 2.0. You bring your own Claude subscription (or API key), so you pay Anthropic for model usage exactly as you would anyway. There's no Neutron license, no seat fee, no entitlement gate.
Not yet. Neutron is pre-release and under active development, the architecture is real and runs, but expect rough edges, breaking changes, and incomplete pieces. Star the repo to be notified when the first stable release lands, and please hold off depending on it until then.
Ryan Junee, engineer-founder with three exits (Omnisio to YouTube, Parsable to CAI Software, Ostro Health to Veeva). Neutron is the productized version of the system he runs his own businesses and life on. Engineering deep-dives at ryanjunee.com/posts.
A neutron is the neutral, stable, fundamental particle. It holds everything together without taking sides. That's the posture: neutral by design, stable under load, built to answer to you and no one else.
One command installs Neutron on your machine. It's pre-release, so expect rough edges, and star the repo to know when it's stable.
curl -fsSL https://neutronagent.ai/install.sh | sh Apache 2.0. Your machine, your data, your Claude subscription. Nothing sent anywhere except the model calls you authorize.