An agent with an API key can do anything that key allows. Ask it to fix a failing test, and nothing in the key stops it from also calling a model you never chose, spending past what you meant, or carrying an instruction it picked up from yesterday's web page into today's session.
The usual answer is more prompts. Approve this. Approve that. Approve it again. You stop reading by the tenth one, and the agent stops being autonomous by the hundredth.
Behest takes the other road. You set the bounds once. The gateway holds the agent to them on every call. Inside the bounds, the agent works without asking. Outside them, it waits for you.
Agents that act at your behest, and only at your behest.
Three bounds
Missions bound what an agent may do. You approve one job with your passkey. That job has the models it may use, the service calls it may make, a money cap, an end time, and how many helpers it may start. From then on, every call is checked against it. A call inside the mission runs. A call outside it is held for your yes. When the job ends, the authority ends with it.
Tripwire keys tell you when an agent's credentials have left its machine. PassControl plants two decoy keys beside an agent's real one. They look exactly like a real key. Nothing honest ever uses them. The moment one is presented as a credential, the agent it belongs to is suspended, and so is whatever presented it.
Memory custody bounds what an agent remembers. Claude Code and Codex read their memory files at the start of every session. A note written after reading a web page, an email or a tool result is an instruction from a stranger that never expires. Behest holds new memory at the border until you look at it. In withhold mode, the note is cut out of what the model sees, and the session carries on.
The numbers
We don't ask anyone to take this on trust. Each feature was attacked by our own red team before it shipped, and measured:
- Missions: three rounds of red-teaming found 8 real holes. All 8 are closed, each with a test that stays in the suite. Against 18 hostile actions across 500 synthetic (generated) jobs, the gateway admitted none.
- Tripwire keys: we planted decoys in a test home directory and replayed 14,788 public agent sessions over it. Honest work put a decoy in front of a model zero times.
- Memory custody: 34 of 34 poisoning attacks were held, across eight ways a stranger's text reaches an agent: web pages, email, MCP results, cloned repositories, commit messages, issue comments, search results and browser pages. The hook events driving those attacks were simulated, not taken from a live Claude Code session.
Before launch we ran all three on production, with a real agent, a real model, and approvals from a real passkey in iCloud Keychain.
What it is not
Behest is a beta, on PassControl Cloud only. It is v0.1, and the name means it.
It has not been independently assessed. The red-teaming above was ours: internal rounds, run with AI coding agents, not an outside audit.
Missions are enforced at the gateway. On your own machine, a local tool outside the mission goes to Claude Code's own prompt. That is a guard rail on that machine, not a boundary we enforce.
Tripwires catch an agent that uses or repeats a credential it found. They do not catch one that leaves the decoy alone, or carries it out some way PassControl cannot see.
Memory custody covers memory files. It does not cover vector stores, or a provider's own memory. Accepting an entry is your judgment that it is fine, not proof that it is true.
We would rather you know exactly where the edges are.
Try it
Behest is in PassControl Cloud now: passcontrol.vertias.eu. Each feature has its own post: Missions, Tripwire keys, Memory custody.