Authenticated delegation
An agent never acts as itself. Every operation is delegated by an authenticated human user and carries that user's authority, never more.
For AI
An AI agent needs its own identity, not yours
In active development
An agent acting through your API key is indistinguishable from you, which means nobody can later say what the agent actually did, or stop it without stopping yourself. Devory Autonomous User Interface, AUI, gives every AI agent a distinct, credentialed identity of its own. It is scoped to what its owner has authorised, revocable on its own, and never confused with the reasoning model doing the thinking behind it.
Why this matters
Discovery instead of regeneration
AI coding assistants write code quickly, but speed does not remove the underlying problem: the same authentication systems, storage layers, notification pipelines and API integrations get generated again on almost every project. Each time, the assistant spends tokens and context solving something that has already been solved elsewhere, then more tokens debugging and refining what it just wrote. The output is rarely wrong so much as it is redundant, a new implementation of something that already existed, built again from nothing.
Devory changes what an agent does with that effort. Instead of generating a payment integration from scratch, an agent can discover a qualified one in the Marketplace, read its interface, and compose it into the application it is building. That does not mean the agent stops writing code. It means the code it writes is the part that is actually specific to the problem at hand, while the solved majority is reused rather than rebuilt. In practice this tends to mean lower token consumption on repetitive implementation work, fewer repeated mistakes, and more consistent results across projects that draw from the same qualified components.
What AUI actually lets an agent do
A structured way to discover approved capabilities, request scoped resources, submit operations and return auditable outcomes, without ever becoming an independent user or an unrestricted system principal.
- Not an unrestricted autonomous agent, and not an independent user
- Not a replacement for human authority: every action traces back to a real user
- Not a general-purpose chatbot
- Not an agent with access to everything in Devory
- Not production-ready or externally connectable without qualification and evidence
The delegation and audit model
External-agent principal
A compliant AI agent is represented as a distinct, typed principal, never indistinguishable from a human clicking a mouse.
Capability discovery
Agents discover approved, catalogued operations by identifier, never a hidden or guessed action.
Resource scopes
Every request is bounded to explicit, granted resources, never open-ended access to everything in Devory.
Operation submission and lifecycle
Operations are submitted, tracked and resolved through a defined lifecycle, not fired and forgotten.
Confirmation and approval
Higher-risk operations require an explicit human confirmation boundary before they proceed.
Audit envelopes
Every operation an agent performs is wrapped in an auditable, traceable record.
Revocation and fail-closed enforcement
Access can be revoked at any time; when a check cannot be made, the operation is denied, not permitted.
Protocol foundations
These are the contracts AUI is built on today. It is not yet an externally reachable, publicly connectable service.
- Capability protocol
- Resource-scope protocol
- Operation protocol
- Operation lifecycle
- Audit-envelope protocol
- External-agent principal foundation
MCP is the transport, not the innovation
MCP (Model Context Protocol) is a common language many AI hosts already speak, including Claude. It carries the request. What actually matters is what it connects to: trusted components, a qualified Marketplace and an accountable agent identity, which is what AUI provides.
An agent calling Devory through MCP still operates as the same distinct, accountable External Agent Principal described above, scoped, revocable and audited, not a generic, anonymous API call. Today, Devory exposes a minimal MCP surface with no live connection to the platform's real capabilities yet. Broader, governed MCP tools are planned, gated on the same qualification discipline as everything else on this page.
Current status
Protocol foundations are implemented: capability discovery, resource-scope and operation contracts, confirmation gating and audit envelopes. No externally reachable network transport exists yet, so external agents cannot currently connect to a Devory deployment over the public internet. External pilot integration is in development and not available now.