One model, two interfaces
People work through web-rendered pages. Agents discover typed tools through the Agent API and MCP adapter. Both operate on the same data and commands.
Principles
The Blueprint is the shared contract.
Agent access is derived from the application itself. There is no parallel tool schema to maintain and no second authorization model to drift out of sync.
Safe by default
Reads are available by default. Mutating operations must be explicitly published and allowed by exact operation ID before an agent can call them.
Deterministic underneath
The runtime does not need an LLM. Every tool call still passes through Zebric's normal authorization, validation, workflow, idempotency, and audit paths.
Agent operating path Discover, constrain, execute, observe
Discover the contract
The running application advertises its Agent API and the operations its Blueprint intentionally exposes.
Generate typed tools
The MCP adapter turns supported read actions, published mutations, and domain commands into bounded tools.
Approve mutations
Opt in by exact operation ID, with an optional human-in-the-loop checkpoint before anything changes.
Execute and observe
Zebric applies the same policies as the web app and records the result with correlation and audit context.
Agent capabilities
Useful work, inside explicit boundaries.
Zebric has first-party agent and MCP packages, agent-facing discovery, domain commands, and examples designed around agent operation.
Query application data
Discover typed collection and record reads with Blueprint access rules applied to the result.
Run domain commands
Offer goal-shaped actions such as assign, approve, escalate, or close instead of exposing raw CRUD edits.
Create and update safely
Generate mutation inputs only from explicitly published actions, omitting protected and non-writable fields.
Require human approval
Pause before a mutation and resume only after a one-time approval or rejection decision.
Validate Blueprints
Use the first-party agent CLI to check a Blueprint deterministically before connecting to a runtime.
Follow live changes
Forward redacted entity and workflow events into supported agent sessions through the application event stream.
Rotate credentials cleanly
Resolve application credentials at request time and keep them out of model-visible inputs, results, and checkpoints.
Trace every operation
Carry run and correlation IDs through generated tools so agent activity can be investigated alongside human activity.
What “agent-native” means here.
Zebric is not a chatbot bolted onto an internal tool. It is an application runtime where human screens and agent tools are projections of the same declared system. The MCP and Agent API surface is currently a technical preview for the Node runtime; Workers parity, interactive chat, and durable interrupted-run recovery are not yet included.