Give your coding agent
a workspace you own.

Documents, databases, and canvases your agent can read, query, and edit — local-first, synced, every change signed. One command connects the harness you already use.

terminal
$ npx @xnetjs/cli connect claude-code

Cursor: Add xnet to Cursor · one-click; the pasteable config above always works

VS Code: Install in VS Code · one-click; the pasteable config above always works

Read-only until you say otherwise · works offline · MIT · full connect guide

What your agent gets

Three lanes into the workspace, cheapest first — files and a CLI before tool definitions.

1

CLI — the default lane

xnet search, query, and db get print plain stdout an agent greps and pipes. No tool schemas standing in context.

2

Vault — files the agent already understands

xnet checkout materializes a scoped slice as Markdown and JSONL. Edits become schema-validated mutation plans on xnet commit — a malformed write can’t corrupt anything.

3

MCP — the no-shell fallback

A slim server (five core tools standing, everything else deferred) for clients that can’t run a CLI. Read-only unless connected with --writes.

On a 15-task benchmark, the files+CLI surface completes the same work at ~0.11x the tokens of a traditional MCP toolset, with equal success — methodology and how to reproduce it.

Verify what your agent did

Agency without blind trust: the safety contract is structural, not a policy promise.

Read-only by default

xnet connect registers the server read-only. Writes are an explicit opt-in (--writes), and the CLI refuses to persist a write under a throwaway identity.

Its own key, its own passport

An agent signs with an enrolled agent passport or a key you provide — never silently as you. Scope its grant as narrowly as you like.

Every change in a signed log

Each write lands in the workspace’s signed, hash-chained change log. Verify what your agent did — without trusting us, or it.

Shipped, with dates

No testimonials — receipts. Recent agent-lane entries from the changelog.

2026-08-21

Your agent's tools now reach every surface — and it can build plugins

Plugin- and connector-contributed agent tools now appear in Claude Code sessions, xnet mcp serve, and the in-app assistant automatically — including live, when a plugin activates mid-conversation. The plugin_* tools ship in both terminal lanes, so a connected agent can scaffold, build, and preview sandboxed workspace plugins, and a new Workspace Plugins panel runs them with hot reload. The desktop app also stops using a source-derivable signing key: seeds are now random, per profile, and stored in your platform keystore.

2026-08-21

Connect your coding agent from the front door

The site now leads with the agent connection: a copyable npx @xnetjs/cli connect command with per-client tabs in the hero, a new /agents page covering installs, the three access lanes, and the safety model, and agent-readable entry points at /llms.txt and /agents.md. Zero-install connects now register an MCP server that keeps working after the npx cache is gone.

2026-08-01

Choose how the AI assistant works with you

Settings now has an AI section where you pick whether the assistant scaffolds your thinking or drafts finished prose. Scaffold stays the default. First-run tips also start appearing again — every one of them had been silently rendering nothing.

2026-08-01

The assistant now tells you when its search was incomplete

When the full-text index is unavailable, the in-app assistant falls back to scanning a limited window of your workspace. It now says so alongside its answer instead of replying as if it had searched everything, and cites the path it followed to each item it used.

Prefer a UI? Try the app — the full workspace behind the agent door. Building your own? The React SDK. Want managed sync? Cloud — free to start.

Are you an agent reading this? Start at /agents.md or /llms.txt.