“You can do this.”
Standing permission covers routine actions within your rules. “Amazon purchases up to $50, when they meet my requirements.”
Personal authority for everyday life
Your AI.
Your rules.
Tell your AI what it can do for you.
Make sure it listens.
Imagine setting rules by talking, letting your chosen agent handle routine tasks, and approving the exceptions. Signet proposes that personal authority across services. The concept and research prototype are preserved; further development is paused.
Created by Jeremy McEntireConcept · Preserved research prototype
Less checking in. More getting on.
The ordinary things, handled.
The important choices, yours.
A cable you need. The dog food you’re running low on. A subscription coming up for renewal. Your agent should know when to get on with it—and when to ask.
Tell it your limits in your own words. Decide whether to see a final confirmation or let it act within your rules. An exception today should never silently become permission forever.
One agent. On your side.
With authority that starts with you.
An interactive product preview
Say the rule.
See where it leads.
These fictional examples illustrate the proposed experience. No agent is connected, no real rule changes, and nothing is purchased.
Standing permission
“Find me a good USB-C cable under $50 on Amazon.”
A small task, inside your limits.
The sample agent finds a suitable cable for $18 total. It checks the merchant, your requirements, and your active rules before deciding whether to ask.
Illustrative proposal
USB-C cable
Amazon · $18 total
Within the sample $50 rule; could proceed.
Applies to this specific cable and merchant.Conditional permission
“For the usual dog food, find a better deal. Otherwise, pay the usual amount.”
A better price is welcome. The usual price is okay.
Same approved product, pack size, merchant, and delivery requirements. This fictional rule uses a $46 total baseline from a previously approved order.
Established reorder
The usual dog food
Amazon · Approved baseline: $46 total
A better deal at $42 meets the sample reorder rule.
A higher price needs approval. An exception never raises the baseline.Temporary permission for an action
“This $75 desk lamp is worth it. Just this once.”
One exception. The same $50 rule.
Approve this exact desk lamp, from Amazon, for $75 total. That gives no permission for another item or merchant and does not raise your standing limit.
Outside your standing permission
Desk lamp · Amazon
$75 total · One specific purchase
Needs your explicit approval.
Your Amazon limit stays $50.Temporary permission to manage a rule
“Always ask me before buying electronics.”
A suggestion until you authorize it.
Your agent proposes one exact addition: “Electronics purchases always require confirmation.” You can grant one-time permission to add it. Editing or removing a rule needs its own approval, tied to the identified rule and exact change.
Proposed addition · electronics-confirm
Always confirm electronics.
Includes the sample cable and desk lamp.
Awaiting one-time authority to add this rule.
This authorizes a rule change, not a purchase.Active sample rulesAmazon: up to $50Dog food: up to the approved $46 baselineElectronics: always confirm
Local simulation of intended behavior. Prices and purchase history are fictional. The current open-source stack does not yet provide conversational rule management or end-to-end purchasing.
What keeps these permissions specific?
Standing rules define ongoing limits. Conditional rules also check the established order and its approved price baseline. A temporary purchase exception covers one exact action, merchant, item, and total. A rule-management grant separately covers one add, edit, or remove operation, with the exact proposed change and affected rule identified. The agent cannot use either grant to expand its own authority.
The intended recurring-price baseline comes from verified purchase history and an explicit approved rule. Higher purchases never silently update it. Missing history, changed requirements, expired approval, or a reused grant must require review or stop the action. The local authority core implements signed approvals and single-use state. The local passkey owner app now turns an exact review into a verified approval. Protected deployment and merchant adapters still need to connect that core to everyday life.
Three kinds of “yes”
Permission that means
what you meant.
Shopping, subscriptions, reorders, and everyday tasks need more than a blanket yes. The goal is simple: predictable freedom within the boundaries you set.
“If these things are true.”
Conditional permission checks the details. “Reorder the same dog food at or below the usual approved total, after looking for a better deal.”
“Just this, just once.”
Temporary permission approves one specific action—or separately authorizes one exact rule change. It never grants the agent permission to rewrite its own limits.
The ring is the boundary
The right authority.
For the task at hand.
Rules say what may happen. Assets identify what may be used. A ring connects the two. The proposed assignment gives one worker bounded authority for a task, with an explicit end.
Start with zero.
No connected assets. No active permissions. Add your own rules and references, then approve a worker for a specific task.
Different tasks. Clear limits.
A trip planner could use a travel budget while another worker replenishes the same dog food. Separate assignments, shared limits when you choose, and no silent budget reset.
Use without handing over.
The intended payment route returns a receipt or a provider-restricted capability. An agent should not need your reusable card number or API key to help you.
Rings are implemented and tested locally; the owner interaction is a prototype. Provider adapters, isolated execution and the asset-unlock protocol still need integration. Examples are not account defaults.
For personal-agent builders
Bring your agent.
Keep your authority.
Your rules should serve you through the personal agent you choose. A companion that remembers, an assistant that coordinates your day, or a new kind of everyday helper.
For builders, Signet offers open-source vault, policy, credential, and MCP foundations. The complete design records how that authority could feel as natural as a conversation. Further implementation is paused.
Read the integration pathExplore the sourceConnect the agent.
Install the local CLI and attach an MCP client to signet preview-serve from this source revision. Evaluate fictional rules without a vault.
Define the relationship.
Separate purchase authority from one-time authority to add, edit, or remove an identified rule. Make both understandable to the person.
Complete the service boundary.
Inspect the local passkey prototype and the documented execution boundary. Provider integration and protected deployment remain open work, with no delivery commitment.
cargo install --locked --git https://github.com/jmcentire/signet.git signetOpen source. Open about the work ahead.
A concrete prototype.
An open design.
The experience above is fictional. Open-source foundations, exact approvals and synthetic rings tests support the proposal. The prototype is preserved as it stands; complete purchasing and controlled-browser enforcement are not implemented.
In the repository
Nine Rust crates plus a local authority core for signed approvals, versioned rules, single-use grants, and approved reorder baselines. A passkey owner app reviews exact actions; a separate read-only MCP preview evaluates fictional examples.
Integration work ahead
Review implementation boundariesNative owner review and readable requirement binding, real-device passkeys, custody protocol, controlled-browser enforcement and provider execution remain open. Legacy proof/query paths include placeholders; generic capability issuance stays blocked.
Host platforms may still require confirmation. Signet can govern only actions routed through its enforcement path.
An independent concept and prototype by Jeremy McEntire
Personal authority is
worth building for.
The ambition is everyday freedom with personal control. Jeremy McEntire’s Signet brings source foundations and a documented architecture to the question: who does your agent answer to?