Native Soil

What is Native Soil?

A continuity layer for AI-assisted work. Save your project's state in one model, load it in any other · without drift.

Save here, load anywhere. Native Soil is a continuity layer for AI-assisted work. When you've built up real context with an AI · a project's architecture, its locked decisions, what's done and what's next · Soil extracts that into a handover another AI can pick up, in a different thread, a different model, or a different provider, without drift.

It's not a note-saver. The hard part, and Soil's whole reason to exist, is doing the extraction smartly: capturing everything a fresh model needs to continue without drift, and nothing it doesn't · never secrets, never raw prompts, never enough to make a model refuse to load it.

Who it's for

Anyone who does serious work with AI assistants and hits the wall where context runs out, a thread gets too long, or you want to move the same project from one model to another. Developers using Claude, ChatGPT, Codex, Cursor, Grok, Zed, Gemini and similar tools are the first audience, but the idea is tool-agnostic by design.

The two ways to use Soil

Honesty rule

Throughout these docs, every feature is badged Available now or Coming. We never document something as live until it is. Both Soil Cloud (the Remote MCP connector) and Soil Desktop (the local, offline-first app) are in beta now.

Quickstart

No account needed: the free template

Want the value without signing up? The free manual handover template is a polished, copy-pasteable prompt that produces a structured handover in any model by hand. It is the same section shape Soil uses, written so a human can use it anywhere. Start there, then graduate to Soil Cloud for the graded, stored, loadable version.

How it works in one breath

  1. In a conversation where you've built up context, you say soil save. Behind that one command, Soil hands the active AI a recipe (a schema + an extraction prompt), the model fills it in from the live conversation, and Soil checks the result: a safety scan, a completeness score, and a drift check ("could a cold model continue with only this?"). Then it stores a safe projection (meaning, never secrets). A complete, high-confidence extraction passes the checks and is saved in one go; a borderline one comes back with the exact gaps to fill.
  2. In a new thread or a different model, you say soil load #001 (an exact code) or soil load #my-thread (by name · name it at save time with soil save #my-thread) and continue where you left off. That bare save is a standalone snapshot, free. When you want work to accumulate, save into a project: each soil save @project adds a thread, and soil load @project brings the project back · the project knowledge in full, every thread listed and attributed, the newest restored inline · so a fresh AI or a teammate picks up the accumulated work, not just one chat. (@ says where, # says which.)

The model in the room does the reading; Soil owns the recipe and the proof. That inversion is what lets Soil work in clients where nothing can scrape your chat history.

Soil's deeper promise is that quality is graded, never assumed: every save comes back checked and graded, and on project loads continuity is measured on your own two models, the one you save from and the one you load into. See how Soil verifies a handover.

On this page