Give each agent one MCP token and it can use the apps you connected, like GitHub, Slack, Gmail, and Notion, plus a memory every agent shares. You choose what each agent may use, and Bezalel keeps a record of what it did.
Claude Code asks to buy printer ink for $38.20. The purchase waits in Bezalel until you approve it, then the agent places the order.Works with
Claude Code
Codex CLI
Cursor
OpenClaw
eve
and any other agent that supports MCP.
Every new agent needs your GitHub, Slack, and Gmail again, and starts with no memory. With Bezalel, it doesn't.
Bezalel holds your team's app connections and one shared memory behind a single MCP endpoint. Connect GitHub, Slack, Gmail, Notion, or any of hundreds of other apps once, and every agent you allow can use them.
Claude Code on a laptop, Codex in CI, and Cursor in the editor each get their own token, so you can see which agent did what and switch any of them off. A new agent next month can use the same apps and memory as the last one.
Apps and memory, plus 5 more
What every agent gets
Connectors
GitHub, Slack, Gmail, Notion, Google Calendar, and hundreds of other apps. Connect each once, then every agent you allow can use it, limited to the actions you allow.
Add the dentist to my calendar.
Memory
One memory every agent shares. What you tell Claude Code, Codex can recall next week. Browse or delete any of it from the dashboard.
Book the same hotel as last time.
Email
A real @goshenemail.com inbox your agents send, reply, and search from, attachments included. New mail can wake an agent instead of waiting for it to check.
Reply to the landlord and cc me.
Texting
Text your agent on iMessage or Telegram and it texts back. Pair your number once so it knows which messages are yours.
Move my 3pm to tomorrow morning.
Money
Save a card once. Agents can shop at supported stores, and every purchase waits for your approval under daily spending limits. Link a bank to track what you spend.
Reorder the printer ink.
A computer
A cloud desktop with a browser. The agent clicks, types, and fills in forms while you watch live from the dashboard.
Find flights to Lisbon under $600.
Sandboxes
Throwaway Linux machines for scripts, builds, and experiments, so untrusted code never runs on your laptop.
Bezalel makes a private key for your first agent and puts it in a ready-to-paste setup prompt on the Agents page.
2
Paste the prompt into your agent
Claude Code, Codex, Cursor, OpenClaw, eve, or any agent that supports MCP. It reads the setup instructions, adds Bezalel to its own settings, and checks what it can reach.
3
Turn on what it can use
New agents start with nothing switched on. Connect your apps, then choose what each agent can use. Your next agent connects the same way and can use the same apps and memory once you switch them on.
Step two, in full
Connect yourself to my Bezalel capability plane: fetch https://bezalel.sh/api/setup and follow the instructions. Your token is <TOKEN>.
And text your agent from anywhere. It does the work on your computer.
Coding agents stop when you leave the keyboard. Run one command next to Claude Code, Codex, Cursor, or OpenCode, and a text from your phone reaches that agent on your machine. It does the job, then replies in the same thread.
Texts wait in a queue while your laptop sleeps. New email and webhooks from other services can start a task too.
A text from your phone asks the agent to move a 3pm meeting to tomorrow morning, and the agent replies in the same thread once it is done.
Each agent gets only what you allow, and Bezalel records what it did.
Bezalel checks permissions and limits on every request, so a confused or tricked agent can't use a capability you didn't turn on, run an app action you turned off, or spend past the limits you set.
An agent's permissions in Bezalel: sandboxes switched on, bank and cards switched off.
Permissions per agent
Each agent sees only the capabilities you turn on for it. For each connected app, you choose which actions any agent may run, so an agent can read GitHub issues without being able to delete a repository.
Approval before money moves
Every purchase waits for your OK and counts against daily spending limits. Email and texts have sending limits too.
A record of every action
The Activity page shows which agent ran an app action, sent an email, or changed a memory, when, and whether it worked. Activity records never include an app action's arguments or results.
Revoke anytime
Revoke an agent and its key stops working immediately. You can also forget any memory or disconnect any app.
