karl.

Download karl

Single archive for Windows x64. Includes the server binary, the VS Code extension, the Ollama bootstrap, and a real proof-of-learning artefact you can inspect. No installer, no admin rights required.

Windows x64 .zip · ~40 MB Linux x64 coming soon

Linux and macOS are coming later. The Karl runtime builds for Linux x64 and for Apple Silicon/Intel, but those archives are not published yet. If you want to be notified when they ship, drop us a line. For best overnight fine-tuning throughput, Linux x64 with an NVIDIA GPU is the recommended target.

I want the full version
Heads up: The demo is the real product with gates. See what’s included and what’s held back below.

File details

Version 0.1.0, released 2026-06-04. Pick the build that matches your OS. Each archive is self-contained: no installer, no admin rights, no registry writes.

Windows x64

Filename
karl-demo-0.1.0-win-x64.zip
Size
41,690,753 bytes (39.8 MB)
SHA-256
a0cb0f539931bf3971c273783da99e789ee2206ff44c5599b56c67a9f93f88da

Linux x64

Coming soon. The Linux build is produced by scripts/build-karl-demo.sh but is not published yet — tell us if you need it and it moves up the queue.

macOS

Not available yet. We're working on a signed, notarised macOS build — subscribe to be notified when it ships.

Verify the download

# Windows
Get-FileHash .\karl-demo-0.1.0-win-x64.zip -Algorithm SHA256

# Linux
sha256sum karl-demo-0.1.0-linux-x64.tar.gz

System requirements

  • Windows 10 (1809+) or Windows 11, x64.
  • Linux with glibc 2.31+ (Ubuntu 20.04+, Debian 11+, Fedora 34+, RHEL 9+), x64 or arm64.
  • macOS support is on the roadmap; not shipping today.
  • ~6 GB free RAM headroom while the model is loaded.
  • ~5 GB free disk for the model (downloaded on first run).
  • VS Code 1.80+ (the bundled extension installs into it).
  • Internet access on first run only, to pull the base model.
  • No admin rights, no installer, no registry writes.
  • For overnight fine-tuning, an NVIDIA GPU on Linux or Windows is recommended.

Quick start

Windows

# 1. Unzip wherever you keep tools
Expand-Archive .\karl-demo-0.1.0-win-x64.zip -DestinationPath .\karl

# 2. Start the server (auto-installs Ollama, pulls the model)
cd karl
.\run-demo.bat

# 3. Install the VS Code extension from the bundled .vsix
code --install-extension .\karl-extension.vsix

# 4. Open any project. Ghost-text completions appear as you type.
# 5. Open the learning dashboard:
start http://localhost:5117/learn

Linux

# 1. Install Ollama if you don't already have it
curl -fsSL https://ollama.com/install.sh | sh

# 2. Extract and launch
tar -xzf karl-demo-0.1.0-linux-x64.tar.gz
cd karl-demo-linux-x64
./run-demo.sh

# 3. Install the VS Code extension from the bundled .vsix
code --install-extension ./karl-extension.vsix

# Optional: install as a user-level systemd service (start on login)
./install-service.sh        # systemctl --user status karl  to check

macOS: not available yet. Ping us if you want it — demand from real users decides the priority order for code-signing work.

First boot takes a couple of minutes while the base model downloads. After that, cold-start is sub-second.

What’s included

  • FIM ghost-text completions in VS Code, on every file Karl can see.
  • Chat panel with workspace-aware retrieval.
  • Agentic plan-and-patch — ask Karl to make a change, watch it patch, run the tests, revert on failure.
  • A real proof-of-learning artefact (base vs. Karl, +5.7 pts) bundled in, with a proof viewer at :5117/proof so you can inspect how promotion is decided.
  • Trajectory capture in the background — reasoning, not just commits.
  • Learning dashboard at :5117/learn — active model, scheduler state, accept-rate.
  • HNSW-backed vector index for sub-second nearest-neighbour search at scale.
  • Single binary, single ZIP, no installer, no telemetry.

What’s held back

  • 14 days of full use, then completions return HTTP 402.
  • 200 inline completions, 30 chat turns, 5 agentic sessions in total.
  • Overnight fine-tuning is disabled. The dashboard shows the loop wired up but the trainer no-ops.
  • Generating a fresh proof on your own repo needs the full version — the trial ships a real recorded artefact you can inspect instead.
  • Cross-team convergence detection requires the full version.
  • Admin dashboard for multi-developer deployment is in the full version only.

If you hit a gate and want past it, see licensing or email us.

License summary

The demo is provided under an evaluation licence. Plain English:

  • You may use it on one machine for evaluation.
  • You may not redistribute it.
  • You may not decompile, repackage, or reverse-engineer it.
  • It comes with no warranty. We will absolutely fix bugs you report.
  • Your code stays on your machine. The demo does not phone home.

The full evaluation licence text is in LICENSE-DEMO.txt inside the ZIP.

Want the full version?

Pilot, design partner, OEM, or a quieter acquisition conversation — the door is open. Fastest path is a one-line email.

contact@karl-ai.se Back to overview