What is Nouride?

Nouride is a program you run on your own machine. It hosts several AI agents — each with its own name, personality and memory — and connects them to Telegram, Discord and WhatsApp so you can talk to them the way you talk to anyone else.

It is one file. It carries its own runtime, so there is nothing to install alongside it: no Node, no Bun, no Python, no package manager. You download it, run one command, and finish the setup in a browser.

What it is not

It is not a model. Nouride does not run an LLM on your machine and does not want a GPU. It calls an endpoint you point it at — Anthropic, OpenAI, OpenRouter, a company gateway, or an Ollama on your LAN. Your RAM is unaffected by which model you choose.

It is not a chatbot builder. There are no flows, no node graphs and no visual editor. An agent is a directory of Markdown files, and the whole of its behaviour is what those files say plus the tools you allow it.

It is not a cloud service. Nothing phones home. The conversations, the memory and the credentials are files in one directory on your machine.

Who it is for

Someone who wants a personal assistant reachable from their phone, on hardware they own.
A small team that wants one bot per role — one for deploys, one for research, one for support — without running one container per bot.
A homelab. A Raspberry Pi 4 with 2 GB of RAM comfortably runs a handful of agents.

The shape of it

   Telegram ─┐
    Discord ─┤                ┌──────────────────────────────────────────┐
   WhatsApp ─┼──── gateways ──┤  one daemon, one SQLite file, one port   │──► your model endpoint
   terminal ─┤                │                                          │
  dashboard ─┘                │  agents · tools · approvals · scheduler  │
                              └──────────────────────────────────────────┘
                                              │
                                     .nouride/  ← config, database, agents, workspace

Everything above happens in one process. Agents are not containers and not subprocesses — an agent is a directory of Markdown that gets assembled into a prompt when a message arrives for it. The only thing that runs beside the daemon is the WhatsApp bridge, and only if you use WhatsApp.

What an agent can do

Out of the box an agent can read and write files, search and fetch the web, run commands, remember things, schedule work, and hand a job to another agent. Every one of those is a named tool you switch on per agent, and the dangerous ones ask you before they run.


Read and write	`read`, `write`, `edit`, `multi_edit`, `glob`, `grep` — inside directories you allow
Run things	`exec` for one-shot commands, `session` for something interactive, background jobs for anything long
The web	`web_search`, `web_fetch`
Remember	`memory_write`, `memory_read`, `memory_recent`, and a curated `MEMORY.md` it maintains itself
Work with others	`delegate` to a subagent, `ask_agent` to another persona, `task_plan` for multi-step work
Learn	`skill_create`, `skill_improve`, `skill_toggle` — all of which produce a draft you approve
Anything else	Model Context Protocol servers, declared per agent

What it costs to run

Measured, not estimated — the provenance for every figure is on the footprint page.


Floor	64 MB of RAM. Verified: 48 MB dies before `ready`, 64 MB reaches it.
Comfortable	512 MB. 1 GB if the agent does real work through `exec`.
Disk	~95 MB installed, plus whatever the agents produce.
CPU	Any. 0.7% of one core at idle — it waits on the network almost all the time.
Needs no	Postgres, Redis, vector database, GPU, or local model.

The daemon links three third-party packages — Hono for HTTP, discord.js and grammY for two of the gateways. Everything else is Bun’s own standard library: the database, the HTTP server, the process spawning, the password hashing.

Install it — Linux, Raspberry Pi, macOS or Docker.
First run — the one-time password, the dashboard, the two questions.
Architecture — if you want to know how before you know whether.