Local-first · Agent-native · MCP-ready

Drop any spreadsheet.
Get a local MCP server your AI agent can query.

Your data never leaves your machine. Obsidian gave us local-first notes. Nexus does the same for structured data.

~/nexus

$npx @pixeldesigns/nexus connect customers.xlsx

# Iris is reading your sheet…Detected 14 columns, 2,308 rowsSubject: SaaS customer pipelineGenerated 12 MCP tools: • list_recent_customers() • find_stale_accounts(days) • top_opportunities() … 9 more

$nexus serve

MCP server: http://localhost:5391/mcpReady for Claude Code, Cursor, or any MCP client.

The problem

Every AI tool wants your data in their cloud.

For most data this is fine. For the data that actually matters (customer lists, financial models, pipelines, the spreadsheets you live in), it's not.

→ Today

Hand your data to a cloud SaaS

Upload the sheet to ChatGPT, Claude Projects, or some AI-for-X tool
Hope they don't train on it. Hope they don't retain it.
Re-upload every time the data changes
Watch your master copy diverge from the version in their cloud

→ With Nexus

The agent comes to your data

Point Nexus at your sheet. It stays right where it is
A local MCP server exposes semantic tools your agent can call
Your master is read-only. Every change is a derivation on top.
Selective cloud publishing only when you explicitly want to share

The four pillars

A local substrate, designed for the MCP era.

Iris, the agent inside Nexus, reads any tabular source, understands it, and exposes a semantically meaningful interface to your AI tooling.

01 / Universal input

Connect to anything tabular

Sheets, CSV, XLSX, SQLite, DuckDB, FileMaker. Iris's semantic understanding is source-agnostic. One CLI command per format.

02 / Agent-native

Semantic MCP tools, not `query(sql)`

Tool names match what your data actually is: find_stale_customers(), low_confidence_decisions(). The agent never sees the underlying data, only tool results.

03 / Non-destructive

Git for spreadsheets

Views, collections, branches, snapshots, annotations, all stored separately. The master is read-only. Experiment freely; revert anything.

04 / Selective publishing

You decide what leaves your machine

Per-derivation opt-in to the cloud. PII auto-redacted. Full audit log of what's shared, with whom, when.

Drop any spreadsheet.Get a local MCP server your AI agent can query.