The anatomy · The gateway
Every request for anything in your Mind passes through one place first. This page covers what that place actually does.
01
One request comes in, whether it's you, someone on your team, or an agent acting for either. It goes through the same gate every time. There's no side door, no second path for requests that seem trusted, no exception for the request that looks urgent. One gate, every request, no matter who's asking or why.
02
The gate doesn't try to work out what you meant. It checks what's been granted against what's being asked for, then answers. No model sits in that decision path, not even a small one bolted on for convenience. The rule that decides is code, not a conversation.
03
Three things.
You can search: find where something lives in your Mind without pulling it back yet.
You can read, and reading has three settings. Ask for just the summary, not the whole thing. Ask for one slice of something long instead of all of it. Or grab a batch of items in a single request rather than fetching them one at a time.
You can list: see what's available in a part of the score before you ask for any of it.
That's the whole surface. Nothing else gets in.
04
Deny by default. If something hasn't been explicitly granted, the gate refuses it, and writes that refusal down, the same as it would an approved read. It doesn't quietly disappear. The Receipts page in this section is what that record actually looks like, and what you can do with it.
05
The gate deciding what a request may see is only half the trip. Something still has to read what came back and write the answer, and that means calling a model. That call also routes through infrastructure you control, not a black box on our end, and you choose whose account it runs on. Point it at a Claude or Codex subscription you already pay for and the call authenticates the same way it would if you'd typed it into that tool yourself. Add a metered API key instead, and it runs through that. Either way, the login is yours, and so is the choice.
I spent years working with a tool that couldn't tell me why it gave the answer it gave. I wasn't going to build a gate I couldn't explain to myself.
ANTONY · FOUNDER
FILED
running since 2026-05 · search, read, and list verbs shipped · subscription and API-key model routing both live · a general remote command-line surface is still being built
NEXT Receipts How you check it →The founding cohort is small on purpose. Each Mind is brought up personally.
MENSARA.AI · THE GATEWAY · MACHINE SURFACE · TEXT/MARKDOWN
# The gateway > Every request for anything in a Mensara Mind passes through one deterministic gateway. No AI model sits inside the gateway or makes the allow/deny decision; it checks permissions against the request instead. The verb surface is search, read (summary, byte-range, or batch), and list. Requests that are not explicitly granted are refused by default, and refusals are recorded, not dropped. Once a request is allowed, answering it still requires calling a model; that call is routed through the same gateway infrastructure, and the owner chooses whether it authenticates through their own Claude or Codex subscription or a metered API key. Canonical: https://mensara.ai/how-it-works/the-gateway ## One gate, every time One request comes in, whether it's you, someone on your team, or an agent acting for either. It goes through the same gate every time. There's no side door, no second path for requests that seem trusted, no exception for the request that looks urgent. One gate, every request, no matter who's asking or why. ## Permissions, not intent The gate doesn't try to work out what you meant. It checks what's been granted against what's being asked for, then answers. No model sits in that decision path, not even a small one bolted on for convenience. The rule that decides is code, not a conversation. ## What you can actually ask for Three things. You can search: find where something lives in your Mind without pulling it back yet. You can read, and reading has three settings. Ask for just the summary, not the whole thing. Ask for one slice of something long instead of all of it. Or grab a batch of items in a single request rather than fetching them one at a time. You can list: see what's available in a part of the score before you ask for any of it. That's the whole surface. Nothing else gets in. ## What happens when the answer is no Deny by default. If something hasn't been explicitly granted, the gate refuses it, and writes that refusal down, the same as it would an approved read. It doesn't quietly disappear. The Receipts page in this section is what that record actually looks like, and what you can do with it. ## Whose subscription answers the question The gate deciding what a request may see is only half the trip. Something still has to read what came back and write the answer, and that means calling a model. That call also routes through infrastructure you control, not a black box on our end, and you choose whose account it runs on. Point it at a Claude or Codex subscription you already pay for and the call authenticates the same way it would if you'd typed it into that tool yourself. Add a metered API key instead, and it runs through that. Either way, the login is yours, and so is the choice. **Status:** running since 2026-05 · search, read, and list verbs shipped · subscription and API-key model routing both live · a general remote command-line surface is still being built ## Prompt ```prompt Run a search against my Mind for a topic I know it holds. Then read just the header or summary of the first result, instead of pulling the whole thing. Show me what came back at each step, and confirm whether the request was allowed or held at the gate. ``` ## Related - [Receipts](https://mensara.ai/how-it-works/receipts) - [Asking](https://mensara.ai/how-it-works/asking)