AgentHub never runs your agent’s code. Your system emits traces and AgentHub receives them. The only model calls AgentHub makes on its own are the ones you ask for: LLM-judge evaluations, Cost Calculator runs, and synthetic data generation.
Who it is for
AgentHub is for admins and AI Builders in your organization. People with the User role (the Viewer database role) can sign in, but every AgentHub page shows “Your account has ModelHub access only” and a Go to ModelHub button. See Roles and permissions.The core loop
- Instrument. Create a tracing key, then have your agent POST its spans to AgentHub’s ingest endpoint as JSON. There is no published SDK yet; ingestion is a REST endpoint. See the tracing quickstart.
- Trace. Each run shows up under Traces with its spans, tokens, latency, cost, and status.
- Evaluate. Score a trace against an evaluator collection. Offline evaluators are deterministic and free; LLM-judge evaluators call a model you have configured.
- Optimize and export. Use the Cost Calculator and Simulations to compare models on the same scenarios, write up results in Reports, and preview reinforcement-learning exports.
What is in the sidebar
The AgentHub sidebar, top to bottom:
Settings sits at the bottom of the sidebar and opens as a window over the page you are on. Only admins and super admins see it. See Settings overview.
The model comparison Playground exists only on CorePillar’s demo deployment. On AgentHub, that slot is the AI Usage dashboard.
Module guides
Overview
The project home page and integration guide.
Model Manager
Provider keys, the Platform Demo Key, and project models.
AI Usage
Spend by person and project, token split, and privacy events.
Cost Calculator
Build an agent, generate scenarios, and run them on real models.
Simulations
Review past Cost Calculator runs and their grades.
Traces
List, filter, and inspect traces span by span.
Evaluation
Evaluators, collections, and runs.
Synthetic data
Turn source documents into an agent spec, records, and scenarios.
Reports
Shareable write-ups of runs, evaluations, and traces.
RL export
Label trajectories and preview a JSONL export.
Synthetic data and RL export are not in the sidebar. Their guides explain how to reach them and how complete they are today.
AgentHub and ModelHub
CorePillar ships two products from one platform. They share the same accounts, organizations, projects, and model configurations.
The models your employees chat with on ModelHub are the ones configured in AgentHub’s Model Manager, and their chat spend appears in AgentHub’s AI Usage dashboard. Membership belongs to the organization, not to a site, so someone invited to your organization can sign in to both. What they can do on each is set by their role. See the ModelHub introduction.
Next steps
Quickstart
From sign-in to your first evaluated trace.
Concepts
Organizations, projects, roles, keys, traces, and evaluators.