Native Soil: a remote MCP server that saves an AI chat's working state and loads it in any other model or client (OAuth 2.1 + DCR)

[andreas-naslund]

Pre-submission Checklist

• This post follows the MCP community guidelines

What would you like to share?

Hello!

Native Soil is a hosted remote MCP server that captures the working state of an AI session (the decisions, the context, the plan) and lets you load it back in any other model, client, or provider. The idea is continuity across boundaries: you hit a context limit, switch from Claude to ChatGPT, or come back tomorrow, and instead of re-explaining everything you run soil load and pick up where you left off.

A few things that might be interesting to this crowd:

• 6 tools, all annotated (readOnlyHint / destructiveHint / idempotentHint / openWorldHint). Only the save tool writes.
• No API key, no install. Remote streamable-HTTP endpoint with OAuth 2.1 + Dynamic Client Registration, so any compliant client runs the browser flow on first use.
• Verification on save, not just storage. A save is not accepted until the extraction passes a verification pass: the server has the model cold-read its own handover and scores drift against the source before the save is committed. A thin or self-contradicting extraction comes back as NEEDS-WORK rather than silently saving garbage.
• Secrets and absolute paths are stripped before anything is stored, and every section must be self-contained (no "see the repo" pointers), because the loading model can open nothing but the handover itself.

Connect:

# Claude Code
claude mcp add --transport http soil https://mcp.nativesoil.dev/mcp/soil

# VS Code
code --add-mcp '{"name":"soil","type":"http","url":"https://mcp.nativesoil.dev/mcp/soil"}'

Cursor (and any client that reads mcp.json):
{ "mcpServers": { "soil": { "url": "https://mcp.nativesoil.dev/mcp/soil" } } }

Example calls an agent makes:

• soil_save captures the current state and returns a load code (#NNN)
• soil_load #42 restores that state cold in a fresh session, any client
• soil_info returns account status (proves the connection is authenticated)

What I'd value feedback on: the save-time verification design. Right now I gate a save on a cold-reread drift score plus a few server-side completeness checks, rather than trusting the model's own confidence. I'd like to hear how others handle "did the model actually capture enough to be useful later" without it becoming either too strict (blocks honest thin saves) or too loose (saves something that looks complete but isn't). Also curious whether collapsing the surface into 6 tools reads as the right granularity.

Thanks!

Relevant Links

Site: https://nativesoil.dev/
Docs: https://docs.nativesoil.dev/
MCP endpoint: https://mcp.nativesoil.dev/mcp/soil
Registry: dev.nativesoil/soil (official MCP registry)