Your agent
Chooses
Models, workflows, and strategies determine what action the agent wants to propose.
What should I do?Financial authorisation for AI agents
Let AI agents use capital without giving them unrestricted control.
Grantline sits between an AI agent's signed intent and execution. The agent proposes what it wants to do; Grantline checks whether its current authority permits it before controlled capital can move.
Your agentChoosesstrategy and intent
Signed proposalAction Planexact intent
Explore outcomesSelect a decision state
Proposed actionRoutine supplier payment$2,500 · signed Action Plan
DecisionALLOW
The proposal may enter the controlled execution path.
Strategy. Authority. Execution.
Grantline owns the middle layer. It does not choose an agent's strategy or pretend that an authorisation result guarantees a downstream transaction will succeed.
Grantline
Authorises
Current Mandate authority and configured checks determine whether the exact proposal may proceed.
Am I allowed to do it?Execution path
Executes
The authorised path reaches controlled capital, while downstream systems can still accept or reject the action.
Did it complete?The agent can change. The authority model stays the same.
Give different agents room to operate inside a defined boundary while the owner keeps control of the capital.
Routine payments
Run ordinary transfers inside a defined boundary, with larger actions routed for review.
Controlled capital movement
Give an agent room to move capital without giving it unrestricted custody.
Automated workflows
Connect recurring workflows to a defined capital pool while owner control remains intact.
Narrower authority
Delegate specialised authority to sub-agents without losing the lineage above them.
Live on testnet
The current X Layer testnet enforces the authority path.
The implementation covers signed proposals, inherited authority, owner escalation, revocation, and committed execution evidence.
- Vault custody and controlled execution
- Typed Action Plans with EIP-712 signatures
- Active Mandates and inherited authority
ALLOW,ESCALATE, andDENYevaluation- Native amount limits and native-balance Preflight
- Owner-approved escalation and re-evaluation
- Delegation, revocation, and replay protection
- Vault and Mandate pausing for temporary suspension
- Validity windows and targeted nonce cancellation
Inspect the enforcementX LAYER TESTNET
View the contractsSee the deployed implementation and manifest.Read how enforcement worksFollow the boundary from agent intent to Vault custody.Inspect testnet evidenceReview the current network and exercised integration.Try the demoTry it now on X Layer testnet.
Committed authority changes, approvals, custody changes, and successful execution events are traceable onchain. A read-only DENY has no state change to record.
The authorisation path
Every action starts as a proposal.
Grantline checks each signed proposal and returns ALLOW, ESCALATE, or DENY; ALLOW permits controlled execution but does not guarantee that the downstream transaction succeeds.
01Intent
The agent decides what it wants to do.
02Signed proposal
The exact intent becomes an Action Plan.
03Authority check
Current Mandate and active lineage are evaluated.
04Decision
ALLOW, ESCALATE, or DENY.
Four concepts form Grantline's core authority model.
Delegation, Preflight, escalation, and revocation add controls around this model while the signing key never becomes unrestricted custody.
Mandate
Defines the authority an agent may exercise against controlled capital.
Action Plan
Captures the exact structured proposal that the agent signs.
Vault
Holds the controlled capital while the owner retains custody and administration.
Decision
Resolves what Grantline permits next: ALLOW, ESCALATE, or DENY.
Before execution, Grantline checks authority and Vault state.
Mandate rules define authority. Preflight checks the projected Vault state. Escalation routes a configured boundary crossing to owner review.
Does the proposal fit the authority granted?
Limits, permitted actions, delegation rights, and current lineage define the hard boundary.
Would the projected Vault balance stay above its reserve?
The current MVP checks the projected native Vault balance and native-USD balance against inherited reserve boundaries.
Does crossing this boundary require owner review?
The owner approves or denies the exact stored proposal, and execution checks current state again.
Authority stays bounded as things change.
Grantline evaluates current authority, preserves the active lineage, and checks an approved escalation again before execution.
New Mandate rules are checked when the proposal is evaluated.
A proposal signed under a wider boundary can stop when the Mandate becomes more restrictive.
Revocation preserves history and stops future use.
Revocation preserves the lineage while preventing an inactive Mandate from authorising new execution.
Approval is checked again before execution.
An approved escalation is checked again before capital moves, so changed authority can still stop it.
Delegation can narrow authority, never expand it.
A sub-agent receives a narrower boundary, while restrictions above it continue to apply.
What that looks like in practice
Effective authority follows the active lineage.
Effective authority is the current Mandate intersected with active ancestor boundaries.
Ownersource of authority
Primary agentbounded operating authority
Sub-agentnarrower execution authority
Where Grantline goes next
Extend the authority model without opening a bypass.
Future work adds policy, external context, integrations, and evidence around the same enforced execution boundary.
Validity and destination controls
Destination and capability policies, and shared authority budgets.
Guardians
Conditions that can bring selected, attributable, and time-bounded external context into authorisation.
SDK and API surfaces
Client tools and sponsored relayer submission around the same underlying contract authority model.
Indexed decision evidence
Indexing and assembled Decision Receipts that connect proposals, authority, approval, and execution.
Grantline
Let agents act. Keep authority bounded.
Give autonomous systems room to operate without turning their signing keys into unrestricted control over capital.