Bezalel

Bezalel is a capability plane for personal AI agents, built by Goshen and currently in alpha. It consolidates agent capabilities — memory, email, money, texting, a cloud desktop, code sandboxes, and third-party app connectors — into a single deployable service accessible over the Model Context Protocol (MCP) via one URL and a bearer token.

What It Is

Bezalel solves a specific problem in the agent ecosystem: every new agent framework requires re-wiring the same capabilities inside a different extension model. Bezalel inverts this by making capabilities the durable asset and the framework the replaceable head. Any MCP-compatible agent runtime connects with exactly two values — a URL and a bearer token — and gains access to the full capability set. Switching frameworks becomes a config change rather than a migration.

Core Capabilities

Bezalel exposes the following capability domains, each controlled by token scopes:

• Memory — Long-term memory shared across every agent the owner runs. Session transcripts are banked automatically, distilled into durable facts in the background, and recalled by any runtime, with browse-and-forget controls in the dashboard.
• Email — A real inbox the agent owns, with send, receive, search, and file operations. Inbound messages become events that can wake an agent rather than waiting to be polled.
• Money — A spend ledger with linked bank data for logging receipts and reconciling real transactions. Virtual cards with server-side limits and owner approvals are listed as coming soon.
• Texting — iMessage integration via a paired number. Agents reply on real threads, and inbound texts wake the right agent through the event router.
• A Computer — A cloud desktop the agent can see and operate (click, type, browse), with a live dashboard view while it works.
• Sandboxes — Disposable code-execution microVMs provisioned on demand and shared across every runtime.
• Connectors — Hundreds of third-party apps (calendars, docs, project trackers, and more) behind managed OAuth, exposed as plain MCP tools.

How the Setup Works

The setup path is designed to be agent-readable end to end. Every page on the site serves markdown from the same URL via Accept: text/markdown or a .md suffix, and llms.txt indexes everything worth fetching. The setup flow is:

1. Sign up and mint a bearer token (looks like bzl_…) in the dashboard.
2. Paste a single prompt into the agent pointing it at the plane; the agent fetches live setup instructions and registers the MCP server in its own harness.
3. All agents share the same plane state — memory, inboxes, money, pairings, and connectors — so switching runtimes requires no migration.

Runtime Compatibility

Bezalel works with any runtime that speaks MCP over Streamable HTTP with a bearer token. The docs provide explicit registration instructions for Claude Code, Codex CLI, Cursor, OpenClaw, and eve, as well as a generic path for any other MCP-compatible harness.

Deployment Model

Each hosted plane serves a single owner with its own database and provider credentials — agents share state with each other but never with other owners' agents. The plane is also self-hostable: it ships as a single container image (Node process, Dockerfile included) deployable to any container host. This single-tenant architecture is a deliberate design choice for personal agent use cases.

Current Status

Bezalel is currently in alpha. The site and API are live, and the one-paste setup flow is functional across supported runtimes. Virtual cards (under the cards scope) are listed as coming soon and currently return "not configured" until launch.