Quickstart · ten minutes, once

Ten minutes to your first AI agent session.

This page is the setup, and you do it one time. You install one small program on a computer that you own. That computer then becomes a machine in your workspace. You start AI agent sessions on it from a web browser, on any device. Your team shares the same workspace, and your files stay on your own computer.

1 · YOU SAY publish this as a link I can send you, in plain language 2 · ON YOUR MACHINE your files, your tools, your agent 3 · IT COMES UP a report, a dashboard, an app 4 · A LINK anyone you send it to with nothing installed

Watch it first under 2½ min

What it is, where it came from, and what it is for. The steps below are the part the film does not cover.

Narrated, two and a half minutes. It opens underground, with no product in it at all.

What the film says, in words

A two-and-a-half-minute introduction to Mycolab. It opens underground, with no product in it: under every forest there is a second forest, one thread is a hypha, enough of them together are a mycelium — one body, no centre — and people call it the wood wide web. Then it turns to your own building, where a workstation you cannot reach and a server awake all night are on the same network and still out of reach, because a shell is not an agent that keeps working while you are in a meeting. So we grew one: Mycolab, one network across the machines you already own. Every machine is a hypha; on any of them you start the agent you already pay for, on your own files; the ground they share is a workspace, the machines you name and the people you let in. A network like this needs a door, and you decide what gets through: a guardian you wrote reads the question and the answer, your files do not cross, and every exchange is recorded. What rises into the light is the fruiting body — a report, a dashboard, a live app, on a link you choose who can open. It closes on where this goes: while you are asleep an agent finishes a fit and one in another lab checks it, every lab on one living ground, and by morning the result is simply there — not emailed, not uploaded. It starts tonight.

Set it up ~10 min, once

Four commands on one computer, then one screen. After that you can stop using the terminal.

You need

  • A computer you can install on and leave running. A laptop is fine. No admin rights, no inbound port — so no firewall rule to ask for. Turn its sleep off if you want overnight runs — on a Mac that is System Settings, Lock Screen. The laptop you watch from can sleep all it likes.
  • Model access for the agent — Claude Code, Codex or pi (a coding agent, not the number). Connect your own account, or take model access on a Mycolab plan. Mycolab drives the vendor’s real tool with whichever one you use.
  • Node 20+, for Codex or pi. Claude Code does not need it.

Not for a batch-scheduled cluster node yet. Mycolab needs a process that keeps running, so it suits a workstation, a login node or a server you control — not a queued node behind Slurm or PBS.

your machinethe connectorone process, dialling outthe meshit never calls you backyou, in any browsera phone, a hotel, anywhereyour firewallone connection out, nothing innothing listens on your machine, so there is no port to ask anyone for

Every machine dials out. That is why connecting a lab workstation needs nothing from your IT department: there is no inbound rule to request, because there is nothing on your side waiting to be connected to.

1 · Make an account

In a browser: hypha.run/signup. Email, a passkey, a six-digit code. No password to choose.

That exact address. Signing up is on hypha.run, not mycolab.hypha.run — the mycolab spelling answers with the app, which looks like a loading screen rather than a mistake.

2 · Connect a computer

One small program dials out and answers calls routed to it. It never listens for incoming connections.

curl -fsSL https://mycolab.hypha.run/install.sh | sh
myco login
myco machine start --daemon

No terminal at all? Mycolab.dmg does this whole step from the menu bar.

irm https://mycolab.hypha.run/install.ps1 | iex
myco login
myco machine start --daemon

Windows warns once — More info, then Run anyway. Not code-signed yet, and we would rather say so. No administrator rights needed.

curl -fsSL https://mycolab.hypha.run/install.sh | sh
myco login
myco machine start --daemon

Statically linked: nothing to install first, nothing to unblock.

Check it worked. myco machine status should list this computer as present. If not, myco login again and repeat — the usual cause is an expired approval.

3 · Install the agent on that machine

This fetches the vendor’s real tool and puts it in your own home directory — no administrator rights, and nothing for apt or brew to do. Mycolab then runs it as a long-lived process, signed in as you. The model access is the part you bring; the program is the part this installs. If it is already there, this updates it.

# one of: claude, codex, pi
myco install claude

claude is self-contained. codex and pi come by npm and need Node 20+ — on an older Node they refuse with a sentence saying so. Changing a machine’s credentials later does not restart the daemon: run myco machine restart after, or the new key is stored and not used, which looks exactly like a wrong key.

4 · Open the app and start a session

Go to mycolab.hypha.run/start and sign in with the account from step 1. Press New session: it lists every computer you have connected and shows which are online. Pick the one your data is on, and choose a folder. The terminal streams into the page — close your laptop and the work carries on, because it is running on that machine and not in this window.

Check it worked. The computer you connected in step 2 is in that list, marked online. That is the whole setup; everything after it is this screen.

Before you leave one running unattended: a session is not sandboxed. It runs as your user, with your files and your tools — that is the point of running it on your own machine, and it is also the risk. Giving someone a session on a machine is giving them a shell on it. myco session spawn <agent> --sandbox runs one in a container instead.

All four as one block to paste
# 1. in a browser: https://hypha.run/signup
curl -fsSL https://mycolab.hypha.run/install.sh | sh
myco login
myco machine start --daemon
myco install claude

On Windows swap the first line for irm https://mycolab.hypha.run/install.ps1 | iex. Skip the last if the agent is already on the machine and signed in. Pointing a machine at a gateway or proxy instead is myco machine auth — see Advanced.

What you do, and what you say

Two different things, and it is worth knowing which is which. You drive Mycolab from a screen; the agent does the work, and that part you ask for in your own words.

You, on a screen

  • Pick the machine. New session lists every computer you have connected and which are online. You choose the lab workstation there.
  • File and track issues — add, assign, comment, close.
  • Run a workflow and read its past runs.
  • Create a channel, and see who may call each one.
  • Turn autopilot on — /autopilot in the composer, and stop it the same way.
  • See what is shared — live links, kept URLs, published work.

These are decisions about your machines, your people and your record, so they are screens you can check later rather than sentences you hope landed.

The agent, in your words

  • “refit every scan in this folder and write me a summary table”
  • “put this up on a link I can send to my co-author”
  • “share the app on port 5173 so Anna can click around it”
  • “keep this URL alive — I have put it in the methods section”
  • “ask the session on the GPU box to re-run the fit and tell this one when it is done”
  • “split this across a few agents, one per sample”
  • “keep an eye on this overnight and message me if it stops”
  • “publish this to the group collection so the others can find it”

You never need the commands behind these — your agent knows them, and Mycolab ships the skills that teach it. Every one of them, with what comes back →

a linkanyone holding the URLyour workspacethe people already in itnamed peoplethe ones you listedno gate at allmust be signed in, and in itmust be one of themwhat you publisheda report, a dashboard, an app

Ask for the one you meant. The default is the top lane — an unguessable hostname, readable by anyone it is forwarded to. That is the right trade for a build report and the wrong one for anything confidential, and an unguessable name is not a confidential one: it is visible to DNS resolvers and in the TLS handshake.

An agent session is not sandboxed. It runs as your user, with your files and your tools — that is the point of running it on your own machine, and it is also the risk. Treat giving someone a session on a machine as equivalent to handing them a shell on it. myco session spawn <agent> --sandbox runs one in a container instead; an outside caller on a channel is not confined by it yet.

Your first real job

Four hundred gigabytes of scans sit on the group’s workstation. You are at home with a laptop. Copying the data is not happening, and neither is driving in.

Start it where the data is

You picked lab-box in the app when you made the session, so it already has that machine’s Python and instrument libraries. Nothing is uploaded. The terminal streams to your browser — close the laptop and it keeps going, because the work is on the workstation, not in the window.

"refit every scan in /data/scans with the Voigt model,
 flag any where reduced chi-squared is above 2,
 and write results.csv"
Give it a stopping condition it cannot argue with

To let it iterate unattended, do not let the model decide when it is finished. Name a command whose exit code decides — your own check script, a residual below a tolerance, “these files exist”. A sentence cannot close the work; a command that exits zero can.

myco session autopilot <id> \
  --goal "refit every scan and write results.csv" \
  --oracle "python check_fits.py"
Send one person the result

A link only your co-author can open, checked against her sign-in — she signs in once at hypha.run, free, and installs nothing. The files stay on the workstation; the link is a door to them, not a copy.

myco serve scans-2026 ./out --access anna@example.edu

Without --access you get an unguessable public link instead — anyone holding the URL can read it, including anyone they forward it to. Right for a build report, wrong for anything unpublished.

Nothing left the workstation except the summary.