Veefore docs

Glossary

Every term Veefore uses with a specific meaning, defined once.

Several of these words have loose everyday meanings and precise meanings here. Where they differ, the precise one is what the system implements.

#A

A2A — Agent-to-agent. A verified connection to an agent outside your organization, invoked through the gateway so the same pipeline applies. →

Action — One protected operation requested by an agent, and the record of how it was decided. Carries full lineage. →

Agent — A software actor with its own cryptographic identity, capabilities and lifecycle state. →

Agent Passport — A signed, versioned, exportable snapshot of an agent's identity, capabilities and trust state. Identity, never authorization. →

Approval — A human decision bound to one specific request. Not a standing permission. →

Assertion — A signed JSON payload proving possession of an agent's private key. Carries a nonce and timestamp, so it cannot be replayed. →

Audience — The value bound into an assertion identifying which gateway it is for. Compared as a string, which is why 127.0.0.1 and localhost are not interchangeable.

Authority — The resolved, effective set of things an agent may currently do, after intersecting every hop of its delegation chain. →

#B

Blast radius — Everything a destructive operation would affect. Previewed before every emergency control, side-effect free. →

Break-glass — Emergency elevated access. Heavily audited, CRITICAL severity, separate permission.

#C

Capability — A named abstract permission an agent holds, like refund.create. Coarser than policy; necessary but not sufficient to call a tool. →

CLIENT tool — A tool Veefore decides and records but your own code performs, inside guard(). What veefore registers. Re-register it as HTTP with a brokered credential to take the key out of your code. →

Control plane — The management API. Organizations, agents, policies, delegations, credentials, audit. People talk to it; agents do not.

Credential broker — The component that seals credential material on the way in and leases short-lived, scoped credentials on the way out — to the tool runtime, never to the agent. →

#D

Delegation — A bounded, time-limited, revocable grant of authority from one principal to another. Can never exceed its parent. →

Destination allowlist — The set of hosts a tool is permitted to contact. Enforced at egress; anything else is DESTINATION_BLOCKED.

Divergence — In OBSERVE, a decision where the effective outcome and shadow.decision disagree. The list of things that will break on promotion. →

#E

Enforcement mode — OBSERVE, SIMULATE or ENFORCE. Lives on the environment. →

Environment — A deployment tier within a workspace. Holds the enforcement mode. DEVELOPMENT, STAGING or PRODUCTION. →

Evidence — A typed, expiring trust record: identity verification, security attestation, compliance evidence and others. →

#F

Fail closed — When a dependency enforcement relies on is unavailable, protected actions are refused rather than permitted. →

#G

Gateway — The enforcement plane. Authenticates agents, runs the pipeline, decides every action. The authorization boundary. →

Generation — A counter on an agent identity, bumped on key rotation. An assertion signed with a superseded key is refused with IDENTITY_GENERATION_STALE.

guard() — The SDK wrapper around a tool function. Asks the gateway first and runs the function only if the call is permitted. →

#H

Hash chain — The linked hashes over audit events that make tampering detectable and localisable. →

High impact — A permission that can never be granted implicitly through a wildcard. Must appear on the role explicitly. →

Hosted Veefore — The Veefore service we run. The CLI connects to it with no address and there is nothing to deploy. →

#I

Idempotency key — A caller-supplied key on a mutating request, so a retry does not duplicate the effect. Reuse with a different body is IDEMPOTENCY_KEY_REUSED. →

Identity — An agent's Ed25519 key pair. Not its name, not a bearer token, not the SDK. →

#L

Lease — A short-lived credential issued by the broker for one action. Statuses ISSUED, USED, EXPIRED, REVOKED.

LIMIT — A policy decision that permits the action in capped form. Sets limits.appliedAmount and velocity caps.

#M

MASK — A policy decision that permits the action with named parameter paths redacted, in the forwarded request and in audit.

MCP — Model Context Protocol. MCP servers are registered, their tools are discovered, and each tool stays unauthorized until a human authorizes it. →

#N

Nonce — A single-use value in an assertion. Reuse is NONCE_REPLAYED, which is what makes a captured assertion worthless.

#P

Pause — A reversible operational stop. No new protected actions. Distinct from suspend, which is the security-driven equivalent.

Pipeline — The twelve ordered stages every protected action passes through. →

Policy — A versioned set of rules over a closed field set, evaluated deterministically on every action. →

Preflight — The checks veefore deploy runs before changing enforcement mode, including replaying recent traffic. →

Principal — Whoever initiated a request: USER, AGENT, SYSTEM or API_KEY.

#R

Registry — The private, organization-scoped directory of agents, for internal discovery. Not a marketplace, and publishing grants no authority.

Revoke — Terminal invalidation. There is no un-revoke.

Risk signal — One of thirteen named observations contributing a weighted amount to a risk score, each with a human-readable explanation. →

Rule set — The JSON artefact that is a policy. Authoritative, stored, and what the visual builder reads and writes.

#S

Scope — Where a policy is bound or a role grant applies: organization, workspace, environment, agent or tool.

Self-hosted — Veefore running on your own infrastructure. Engineers name its address to the CLI once. →

Shadow decision — What enforcement would have decided, recorded in OBSERVE and SIMULATE alongside the effective decision.

Step-up — A fresh re-authentication required at the moment of a high-impact operation, regardless of session age. →

Strictness — The precedence order that resolves multiple matching rules. BLOCK > REQUIRE_APPROVAL > STEP_UP_AUTH > LIMIT > MASK > ALLOW.

Suspend — A reversible security stop. Same effect as pause, different intent and audit severity.

#T

Tool — A registered, callable thing with an endpoint, a required capability, a destination allowlist and a risk rating. →

Total — A property of the policy evaluator: every operator is defined for every input, and an unresolvable comparison is false rather than an exception.

Trace id — A correlation id on every response, action record, audit event and log line. Paired with request_id.

#W

Workspace — A team or product area within an organization. Divides resources; does not divide members or roles. →

Identity, authority, policy and enforcement for autonomous AI agents.