For engineers
A governed call is the only call worth making.
Ungoverned MCP is the plain HTTP of the agent era: it works, right up until a counterparty asks you to prove what your agent was allowed to do. ALEETH Authority is the live gate: 51 tools at https://mcp.aleeth.com/mcp, OAuth 2.0, fail closed. Every consequential tool call is evaluated before it runs and sealed after. You do not ban the ungoverned call. You make the governed one the only handshake a bank, an insurer, or an auditor will accept.
The contract
No receipt, no tool.
One rule, and the gate is beneath the agent, so there is nothing to talk it out of. Every consequential call is a transaction the gate has to authorize first.
// Every consequential tool call is evaluated before it runs.
gate.evaluate(
agent = ed25519 identity // who is acting
purpose = the goal it may optimize // why
tool = "payments.wire_transfer" // what
args = [ ... ] // on whom, with what
)
// The gate returns a single-use, host-bound capability, or nothing.
=> ALLOW cap_9f2c1a7b the call proceeds, and it leaves a receipt
=> DENY (no capability) the call never leaves
// Governance failure is a refusal, not a silent pass. Fail closed.The install
One host. One tool class. One day.
You do not adopt a platform to try this. You put a gate in front of one thing and watch it produce proof.
- 01
Point one host at the gate
Put the ALEETH gateway in front of a single MCP host or agent runtime. It is a gate in front of your stack, not a rewrite of it. Nothing about your tools or your models changes.
- 02
Take the read-only evaluation credentials
We hand them over up front, no sales call in the way. Your agents start making governed calls against the gate in a sandbox the same day.
- 03
Gate one consequential tool class
Pick one blast class that can hurt, payments, production mutation, or an irreversible delete. Every call in that class now has to clear the gate before it runs.
- 04
Read the receipt
Each governed call leaves a signed, tamper-evident receipt you can verify offline, with no call back to us. That is the whole proof, produced in your own environment.
No lock-in
A gate in front, not a fork of your stack.
Keep your models, your tools, your MCP hosts, your agent framework. ALEETH does not replace any of it. It mediates the consequential calls and leaves the rest alone. Remove the gate and your stack is exactly where it was, minus the proof.
What the gate adds
A signed receipt, a live halt, and containment.
Every governed call is sealed to a receipt anyone can verify offline. A live agent crossing the line is halted and quarantined. A contained agent has no network path out. All of it beneath the agent, none of it inside the model where a jailbreak could remove it.
Start
Gate a tool call before you finish your coffee.
Read-only evaluation credentials for the gateway are yours on request, no demo wall. Bring one MCP host and one consequential tool class, and leave with a receipt you verified yourself.