MentatLab
MentatLab is Mission Control for DAG-based agent workflows in private environments. It is designed around direct model/API nodes, internal services, and observable execution plans, and it is the workflow layer of this stack: it composes private models and governed tools into graphs an operator can inspect.
MentatLab, on the record
What this site can read about MentatLab, and what it cannot.
- Docs
- Factory
- Telemetry
DAG orchestration over direct model calls
Define nodes and edges, inspect the execution plan, then follow run events in the operations console.
Design
DAG-first design
Flows are explicit graphs, not hidden chat transcripts. The orchestrator design describes validating each graph, rejecting cycles, and planning dependency order so parallel branches show up in the plan.
Runtime
Direct model/API nodes
The intended runtime shape is direct API model calls and internal services. In a private deployment, those model calls can point at FlexInfer endpoints.
Boundary
Context stays governed
Workflow nodes reach tools and context through Loom. MentatLab consumes that boundary instead of replacing it.
Visibility
Operator visibility
Mission Control gives teams a place to inspect definitions, plans, run state, and events when workflows move from experiments to operations. The architecture notes describe a flight recorder that keeps runs, checkpoints, and console events.
Where MentatLab fits
Mission Control interface for building, validating, monitoring, and executing DAG-based agent workflows in private environments.
Good fit
Repeatable, observable workflows
Workflows an operator runs again and again and needs to see: a graph, direct model/API nodes, observable runs, and controlled private infrastructure.
Not the center
Not a coding-harness replacement
MentatLab is not meant to replace a frontier coding harness. It should call models and tools through explicit workflow nodes, not depend on opaque coding-agent semantics.
Stack relationship
Private models, tool access, and workflow control
MentatLab composes what the other products provide into workflow graphs: it is the DAG design and run-visibility layer over direct API model calls and internal services. Each product keeps its own boundary.
- FlexInfer
- Hosts private model endpoints and customization paths that workflow nodes can call through direct APIs.
- Loom
- Governs MCP, context, agent fleet state, sandbox execution, RBAC, and audit around workflow execution.
- fi-fhir
- Provides a concrete healthcare ETL workload where graph orchestration and private AI assistance have a real data boundary.