Why can't anyone open my ClearML task link?

Because a self-hosted ClearML server usually lives on an internal network. The link works on the VPN and resolves to nothing outside it — which is most of the people who ask about your results.

How do I share ClearML results externally?

Pull the scalars, plots and artifacts you need from the task via the SDK, render them into a page, and publish that. The server stays internal; the conclusion travels.

Does this weaken the isolation we self-host for?

Only if you publish the wrong things. You choose what goes on the page — a metrics table and three plots, not the dataset — and the report can be access-restricted.

Can this run automatically at the end of training?

Yes. Add a publish step at the end of the task so every completed run updates the same report, and the link in your tracking ticket is never stale.

What about ClearML's own report feature?

It shares the server's reachability. A report on an internal server is still only openable from inside the network.

Share a ClearML experiment

ClearML is often chosen precisely because it can be self-hosted: the server runs in your VPC, the data never leaves, and the compliance conversation is short.

That same property is why nobody can see your results. http://clearml.internal:8080/projects/… is a perfect link on the VPN and a connection timeout everywhere else.

Self-hosting is the feature and the problem

The people who need to see a model's results are rarely all on the network:

A product owner working from a laptop with no VPN client.
An external reviewer or auditor.
A collaborator at a partner org.
Your own team, on a phone, in a meeting room.

Handing them a VPN and a ClearML account to see one metrics table is a disproportionate ask, and the usual fallback — screenshotting the scalars tab into Slack — throws away every number's precision. See stop screenshotting reports.

Pull the results, publish the page

The SDK gives you everything the UI shows:

from clearml import Task

task = Task.get_task(task_id=task_id)
scalars = task.get_last_scalar_metrics()
params  = task.get_parameters()
# render a page from scalars, params and the plots you care about, then publish

Publish one document containing the comparison against the incumbent, the two or three plots that carry the argument, the parameter set, and the task id for anyone who is on the VPN and wants the raw run.

The task id in the footer matters: it keeps the internal system of record authoritative while letting the conclusion travel.

You choose what crosses the boundary

The objection to any external publishing in a self-hosted shop is, correctly, "what leaves?"

The answer here is: exactly the fields you put in the page. A metrics table, rendered plots, hyperparameters. Not the dataset, not the model weights, not the artifacts. And the report itself can be access-restricted rather than open — see security and the subprocessor list if procurement asks.

Automate it at the end of the task

task = Task.init(project_name="churn", task_name="v31")
...
publish_report(title=f"churn · {task.name}", html=render(task))

Every completed run refreshes the same report id, so the link in your tracking ticket always shows the current state, with the previous round's comments still attached. If the schedule is the hard part, routines can own it.

Worth knowing

Static plots, not iframes. Embedding the internal UI produces a page that fails for exactly the audience you built it for.
Publish numbers as text, not pictures. A screenshotted metrics table cannot be searched, copied or diffed.
Comments anchor to the figure — see commenting on HTML.
5 MB per report body.

Try it

Free — unlimited reports, commenters and revisions.

Read the API reference →