Roster for Anthropic
Use Claude models to interpret who should act.
Configure Anthropic as Roster's resolver model provider for approval routing, workflow ownership, human review, escalation, and delegation-aware participant resolution.
Your deployment controls the Anthropic credential, selected Claude model, endpoint, data-handling arrangement, and production validation.
Last verified 07/13/2026
Claude interprets organizational intent. Roster supplies the governed context.
Roster maintains the operational source of truth for:
“Who should handle the Atlas security review while the primary owner is unavailable?”
Claude helps interpret the request. Roster then validates the model's participant selection and resolves it into active users, groups, roles, agents, or delegated contacts.
- Projects
- Participants
- Users and groups
- Directory membership
- Labels and metadata
- Roles and responsibilities
- Active delegations
- Authorization boundaries
Anthropic does not become the organizational database. Roster remains the participant-resolution and governance layer.
How the integration works
Roster calls Anthropic's Messages API with structured schema, then validates the model's participant selection against its own authorization and directory context.
Agent, workflow, or application
↓
Roster Resolve
↓
Authorized participant context
↓
Anthropic Messages API
↓
Structured participant selection
↓
Active users, groups, roles, or delegatesWhy use Anthropic with Roster?
Configure Anthropic
Roster uses the Anthropic Messages API when ROSTER_MODEL_PROVIDER=anthropic.
ANTHROPIC_API_KEY=<anthropic-api-key>
ANTHROPIC_BASE_URL=https://api.anthropic.com
ROSTER_MODEL_PROVIDER=anthropic
ROSTER_MODEL_NAME=claude-opus-4-8ANTHROPIC_BASE_URL is optional when using the standard Anthropic endpoint. Set it only when using an approved proxy, regional endpoint, or supported provider intermediary.
Store the API key in the deployment secret manager or protected runtime environment.
Quality-first — recommended for complex resolution questions:
ROSTER_MODEL_NAME=claude-opus-4-8Full Resolve option — evaluate for latency, cost, availability, and data-retention fit:
ROSTER_MODEL_NAME=claude-fable-5At the time of verification, Anthropic documents a specific 30-day retention requirement for Claude Fable 5 and states that the model is not available under a Zero Data Retention arrangement. Confirm the current requirement before selecting it.
Reasoning-effort behavior — direct Anthropic requests currently ignore ROSTER_MODEL_EFFORT even though Roster accepts the variable globally. Validate model behavior through model selection, Roster's prompt and schema, output-token limits, representative Resolve evaluations, participant-model quality, metadata design, and safe fallback rules.
Example resolution
A customer-operations agent asks:
“Who should receive the Acme escalation for Canada, given that the normal enterprise account owner is away?”
- Roster identifies the accessible Acme project.
- Roster collects relevant escalation participants.
- Roster includes region, customer segment, membership, and delegation context.
- Roster sends the structured interpretation task to Claude.
- Roster validates the selected participant.
- Roster expands it into active users, groups, or delegated contacts.
- Roster records the Resolve request and model invocation.
The calling system can then create a ticket, send a message, or assign the case.
Structured output and observability
Roster depends on predictable structured output, not merely fluent prose. The selected Claude model must reliably return parseable data, follow the expected schema, reference valid participant identifiers, avoid inventing unavailable participants, represent no-match cases correctly, and remain stable across similar queries.
Anthropic structured outputs can constrain compatible Claude responses to a defined format. Roster still validates the result against its own participant and authorization context.
Roster records provider and model, status, live or test mode, start and completion time, latency, input/output/total tokens, estimated cost, Resolve request and trace IDs, provider request ID, input and output payloads (subject to PII settings), and errors. Use Model Runs to compare the Anthropic configuration with other approved providers using the same evaluation cases.