Bezalel: a capability plane for personal agents

What we're building

Bezalel answers a question every serious agent deployment eventually hits: where do an agent's capabilities actually live? Give each agent its own email access, payment rails, and memory, and you get a sprawl of credentials and integrations nobody can audit.

Bezalel is a capability plane: one deployable service that owns the capabilities (memory, email, money, a computer, and more) and serves them to any agent runtime over MCP, the open protocol that Claude, Cursor, and most modern agent tooling already speak. The agent doesn't hold the keys. The plane does, and every runtime consumes the same governed surface.

How it's built

• An Effect-based service exposing two surfaces: the MCP surface agents consume, and an admin surface for operating the plane, with typed contracts and state on Postgres.
• Consent built into the flow. Sensitive capabilities go through explicit consent handoffs, tested by their own suite, so a human approves what the agent is about to do with real accounts and real money.
• Real money, done carefully. Financial capability syncs through Plaid, with its own dedicated test suite in the contract matrix.
• A dev console with live status, token minting, and an MCP playground, so operators can inspect exactly what the plane is serving.
• Tested like infrastructure. A contract test matrix runs on every push, and end-to-end canaries exercise the whole plane daily. Coding agents (Cursor, Claude Code) are live consumers of the plane during development, so the MCP surface is dogfooded constantly.
• Packaged for consumers as an eve framework extension and a skills sync tool, so an agent picks up the plane's capabilities and its usage playbooks together.

Why it matters

Capability governance is the difference between an agent demo and an agent you can trust with your inbox and your money. Bezalel is our answer in the open: one auditable place where capabilities live, explicit consent on the sensitive ones, and infrastructure-grade testing. It's the same architecture thinking we bring to client agent deployments.