Most tools advertise local models and then quietly disable half the product when you turn one on. Agent mode goes away. Long context goes away. The thing that made the tool worth using turns out to have lived on someone else's servers all along.
Offline is the reference, not the fallback
We build the offline path first and treat the hosted path as an optimisation of it. If a capability cannot be expressed against a model running on a laptop, it does not get built into the core — it gets built as an explicitly optional layer that announces itself. That constraint has killed a few features. It has also meant that the feature set does not silently shrink when you disconnect.
- Repository indexing runs locally and incrementally; it never uploads a tree to be embedded elsewhere.
- Agent planning is model-agnostic — the plan is a data structure, not a prompt transcript.
- Context assembly is budget-aware, so a 32k local model gets a different, smaller, still-correct context than a frontier one.
- Verification runs against your real test command, which never needed a model in the first place.
What genuinely costs you
We would rather be specific than reassuring. A local 30B-class model is slower per token and weaker at multi-file reasoning than a frontier model. Long refactors take more passes. What does not change is the shape of the work: the same review, the same verification, the same rollback. You trade throughput, not correctness — and no prompt, no file, and no telemetry leaves the machine.
Privacy that costs you the feature you came for is not a privacy feature. It is a downgrade with better marketing.
Klarcode engineering notes