karl.

The code assistant that re-trains itself overnight.

Karl runs on your machine. It learns from the code you write, fine-tunes a candidate on a schedule, and swaps in the new model the moment it beats the old one. No cloud round-trip. No telemetry. No per-seat licence.

Why now

The tech is ready

2026 proved AI can really code

A year ago, “train your own coding model” was a science project. This spring, frontier models showed the category is real — and that a smaller model tuned on your codebase can punch far above its size on the code your team actually writes.

The incumbent blinked

Per-seat pricing just jumped

The cloud assistants raised their per-seat prices, and whole engineering orgs are openly looking for an alternative they own outright. Karl is that alternative: a one-time licence, runs on hardware you already have, no per-seat meter.

The gap is open

Nobody owns “your own model”

The few who tried self-hosted fine-tuning pivoted to cloud agents and walked away from it — just before the price shift made owning your model matter. Karl walks in through the door they left open, with the one thing they never shipped: proof the fine-tune improved on your code.

How it works

Local-first

Your code stays on your machine

A single Windows binary plus a VS Code extension. The model runs against a local Ollama backend. No outbound traffic, no analytics. Works offline once the model is pulled.

Self-learning

Gets sharper on your repo

Karl mines fill-in-the-middle training pairs from your commits and from the suggestions you accept. A LoRA fine-tune runs on a schedule. The candidate only goes live if it beats the incumbent on a held-out benchmark. On a team, one GPU box can train and the rest pull the promoted model automatically.

Trajectory-aware

Captures how your team reasons

Karl indexes the full path your team takes to solve a problem — not just the final commit. That powers duplicate-work detection across teams, “already decided” guardrails, and search that finds the person who once worked through the same reasoning.

By the numbers

+3.2 pts benchmark uplift after one overnight LoRA cycle — candidate vs. base, 266 pairs, 71 holes.
~98% recall@10 on the HNSW trajectory index at 200k vectors, sub-second.
0 bytes of telemetry. No outbound traffic after the model pull on first run.

Base model qwen2.5-coder:1.5b-base scored 0.406. Karl’s candidate scored 0.438 after one cycle. First-line exact-match jumped from 0.352 to 0.408 (+5.6 pp). Promotion rule: the candidate has to beat the paired baseline on the same seed and the same holes by at least a full point, otherwise it’s discarded.

Every cycle writes a JSON evidence artefact to disk and renders it in Karl’s built-in proof viewer. The free trial ships a real recorded artefact from one of these cycles — open the proof viewer and inspect exactly how the promotion decision was made. The full version runs the cycle against your own repo and writes a fresh artefact every night.

Quick start

Three commands. Five minutes.

# 1. Download and unzip
Expand-Archive .\karl-demo-0.1.0-win-x64.zip -DestinationPath .\karl

# 2. Run the demo (auto-installs Ollama if needed)
cd karl
.\run-demo.bat

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

Open any project in VS Code — ghost-text completions appear as you type. The learning dashboard lives at http://localhost:5117/learn.

Get in touch

The demo is a taste. The full version is the product.

The download is a 14-day evaluation: 200 inline completions, 30 chat turns, 5 agentic plan-and-patch sessions, fine-tuning switched off. Enough to feel how Karl behaves on real code.

The full version lifts the gates, runs the overnight fine-tuner against your repo, ships the trajectory and convergence layer, and includes the admin dashboard plus on-prem install. Pilots, design partners, and quieter acquisition conversations all welcome.