Lumo Copilot v0.2.5

Lumo Copilot v0.2.5 — Lumo now refuses cloud models.

Why this matters
Ollama has started listing hosted models alongside local ones — Kimi K3, for
example, is kimi-k3:cloud: 2.8 trillion parameters, billed per token, running
on someone else's servers. They appear in the same registry as models that run
on your Mac.

That was a hole in Lumo. The "Add any Ollama model by name" field would have
accepted one, and you could have started sending your conversations off your
machine while the app still told you nothing ever leaves it.

Lumo now detects hosted models and refuses them, with a plain explanation:

That model runs on remote servers, not on this Mac. Lumo only uses local
models so your conversations never leave the machine.

Checked at both download and selection, so one can't slip in even if it was
installed outside the app.

To be clear about Kimi K3 specifically: it isn't coming to Lumo. Beyond
being hosted, 2.8 trillion parameters is roughly 1.4 TB of memory — it could
never run on a Mac regardless.

Everything else is unchanged from v0.2.4.

Lumo Copilot v0.2.4

Lumo Copilot v0.2.4 — proper installer window, and fewer security prompts.

The install window looks right again
v0.2.3 fixed the "damaged" error but built the DMG by hand, which lost the
styled drag-to-Applications window. Signing is now part of the normal build, so
you get the correct installer and a valid signature.

About the security warnings after v0.2.3
macOS ties an app's granted permissions to its code signature. v0.2.3 was the
first build with a real signature, so macOS treated Lumo as a brand-new app and
asked for everything again. That was a one-time consequence of the fix.

The app's identity is now stable (com.lumocopilot.app), so permissions stick
from here. You may still get a prompt after an update — unavoidable without a
paid Apple certificate — but not the flood you just saw.

First launch, if you're installing fresh
Double-click, let macOS refuse it, then System Settings → Privacy & Security →
Open Anyway. Once only.

No functional changes — same app as v0.2.2, correctly packaged.

Lumo Copilot v0.2.3 — fixes "damaged" error

Lumo Copilot v0.2.3 — fixes "Lumo Copilot is damaged and can't be opened".

If you hit that error, this release fixes it. Nothing was wrong with your
download — every build before this one shipped with a broken code signature, and
macOS refuses those outright on Apple Silicon with a misleading "damaged"
message. The app bundle is now correctly signed, so it opens normally.

(Cause, for the curious: the project folder is synced by iCloud, which stamps
every file with extended attributes. Those quietly corrupted the signature at
build time. Builds are now signed in a clean location.)

Opening it the first time (macOS 15 and newer)
Apple removed the old right-click → Open shortcut. Now:

1. Double-click the app — macOS will refuse it.
2. Open System Settings → Privacy & Security, scroll to Security.
3. Click Open Anyway next to the Lumo Copilot message, then confirm.

You only do this once. If you'd rather use the terminal:
xattr -dr com.apple.quarantine "/Applications/Lumo Copilot.app"

This is the trade-off of a free, unsigned app — Apple charges for the
certificate that removes the prompt.

Everything else is unchanged from v0.2.2.

Lumo Copilot v0.2.2

Lumo Copilot v0.2.2 — the Coding editor gets its space back.

The terminal no longer takes 40% of the screen doing nothing
The Terminal / Diff / Preview panel was claiming a fixed 42% of the column
height whether or not it had anything in it — so an empty terminal was crowding
out the file you were actually trying to read.

It now starts collapsed to just its tab bar and opens itself the moment it has a
reason to: when you run a command, when you click a tab, or when you open a diff.
Nothing is hidden and nothing is harder to reach — the editor simply gets the
room by default.

Also in this release: verification, not just features
After a run of feature releases, this one re-tested the fundamentals end to end
against a real model — a full chat turn with a real tool call, the tool
executing, and the model correctly continuing from the result. All endpoints
healthy, clean startup, no regressions from the last several updates.

Install / update
Existing installs update automatically. New installs: right-click the app → Open
the first time (unsigned build), Apple Silicon only.

Lumo Copilot v0.2.1

Lumo Copilot v0.2.1 — a calmer file tree in Coding.

Build output is hidden by default
The Coding file tree was showing build/, dist/, release/ and stray bundled
files right alongside your actual source — folders you never edit, sitting at the
top of the list burying the code you do. They're hidden now, with a small
"Show build output" link at the bottom of the tree when there's something
hidden, so nothing is out of reach.

Also covers out, target, coverage, vendor, .next, __pycache__,
venv, lockfiles and sourcemaps.

Note this only affects what you see. Lumo itself can still read those folders
when a task genuinely calls for it.

Files are colour-coded by type
Code, config, styles and docs each get their own tint, so the tree can be scanned
at a glance instead of read line by line.

Clearer empty state
The editor pane now says something useful: pick a file, or just describe the
change and let Lumo find the file itself.

Install / update
Existing installs update automatically. New installs: right-click the app → Open
the first time (unsigned build), Apple Silicon only.

Lumo Copilot v0.2.0

Lumo Copilot v0.2.0 — Deep Research and Compare.

Two new tools live under Research in the sidebar.

Deep Research
Ask a real question and Lumo does the legwork: it breaks the question into
searches, runs them, actually opens and reads the pages it finds, then writes one
answer with inline citations back to each source. You watch it work — every
search and every page it opens is listed as it goes — and the sources are linked
at the bottom so you can check its work.

Needs Internet access enabled (Settings → Permissions). Everything else still
happens on your Mac; only the pages it fetches come from the web.

Compare
Pick up to four of your installed models, ask them all the same thing, and read
the answers side by side with timings. This is the fastest way to find out which
of your local models is actually good at the kind of work you do — rather than
guessing from parameter counts.

Models run one at a time on purpose: they share your Mac's memory, so running
them at once would just thrash.

Install / update
Existing installs update automatically. New installs: right-click the app → Open
the first time (unsigned build), Apple Silicon only.

Lumo Copilot v0.1.10

Lumo Copilot v0.1.10 — fixes models silently losing their tools.

The bug
If Ollama was still starting up when Lumo launched — or if you pulled a new model
after the app was already open — Lumo could check that model's capabilities once,
get no answer, and then remember the non-answer. The model would run chat-only
for the rest of the session: no reading files, no running commands, no
explanation. Restarting the app was the only fix.

Lumo now only remembers a real answer and retries otherwise, so a model that
supports tools gets them as soon as Ollama is actually reachable.

This was found while testing something unrelated: a model that Lumo reported as
"chat-only" turned out to be fully tool-capable when asked directly.

Install / update
Existing installs update automatically. New installs: right-click the app → Open
the first time (unsigned build), Apple Silicon only.

Lumo Copilot v0.1.9

Lumo Copilot v0.1.9 — one "thinking" indicator, not two.

Fixed
v0.1.2 added the live thinking view, and it showed a second "Thinking…" row
alongside the existing " is thinking…" indicator — two of them on screen
at once, the first one empty.

Now there's always exactly one:

• While the model is starting up or working quietly, you get the familiar
  " is thinking…" line.
• The moment it actually streams its reasoning, that's replaced by the
  expandable thinking block showing the real chain-of-thought — click to expand
  or collapse it.
• Once the answer starts, the block collapses to a one-line
  "Thought for Ns" summary.

Models that don't expose their reasoning simply show the plain indicator, as
before.

Install / update
Existing installs update automatically. New installs: right-click the app → Open
the first time (unsigned build), Apple Silicon only.

Lumo Copilot v0.1.8

Lumo Copilot v0.1.8 — Lumo can use external tools now (MCP).

What's new
Lumo speaks the Model Context Protocol — the same open standard Claude Code and
other tools use — so it can borrow tools from external MCP servers instead of
being limited to its six built-ins.

Add one under Settings → Permissions → Tool servers (MCP). For example, the
official filesystem server:

• Name: files
• Command: npx
• Arguments: -y @modelcontextprotocol/server-filesystem ~/Documents

Servers start with the app, and Lumo shows how many tools each one offers.

Three deliberate limits
This is the one feature that can reach outside your machine, so it's built
carefully:

1. You add every server yourself. Nothing is auto-discovered or enabled by
   default.
2. Every MCP call still asks you first — the same approval card as running a
   shell command, showing which tool and what arguments.
3. Only capable models get them. Testing showed smaller local models already
   mishandle six tools, and a longer list makes them worse — so MCP tools go only
   to models that handle tools well, capped at 12 at a time.

A word of caution: an MCP server is a separate program running with your
permissions, and many of them use the internet. Only add ones you trust. Lumo's
own chat still runs entirely on your Mac.

Install / update
Existing installs update automatically. New installs: right-click the app → Open
the first time (unsigned build), Apple Silicon only.

Lumo Copilot v0.1.7

Lumo Copilot v0.1.7 — commands no longer flood the model's memory.

The problem
Local models have small context windows (often ~8k tokens), and nothing fills one
faster than a noisy command — an npm install, a test run, a big find. Worse,
Lumo used to keep only the first 4,000 characters of output. For a build or a
test run that's exactly backwards: the error is at the end, and it was being
thrown away before the model ever saw it.

What's new
Tool output is now compressed before it reaches the model:

• Colour codes and progress-bar redraws are stripped — pure noise.
• Runs of identical lines collapse to one, tagged with the count
  ([× 300 identical lines]), because "this repeated 300 times" is itself useful.
• When output is still too long, Lumo keeps both ends and marks what it
  dropped, favouring the tail where errors live.

On real commands this saved 33–93% of the space, and in the two largest cases the
final line — previously discarded — is now preserved.

Files and directory listings get the same treatment, tuned differently: files
favour the top (imports and definitions orient the model), long listings keep
every filename but drop per-entry metadata.

Your view is unchanged. Compression applies only to what the model reads —
the approval card still shows the real, full output, and now notes when a
result was trimmed.

Install / update
Existing installs update automatically. New installs: right-click the app → Open
the first time (unsigned build), Apple Silicon only.