Roster for MCP and custom AI agents

Add a reusable “who should act?” capability to any compatible AI agent.

Roster exposes participant resolution through a Streamable HTTP MCP server, giving compatible agents a standardized way to resolve approvers, owners, reviewers, groups, roles, service accounts, and delegated contacts.

Build the agent once. Keep organizational routing in Roster.

Connect an MCP clientExplore the MCP tools

Last verified 07/13/2026

Agents need organizational context — not another hard-coded routing table

Agent frameworks are good at deciding which action to perform, which tool to call, which data to retrieve, and which workflow step comes next.

But production workflows also require a separate decision:

Who should own, approve, review, receive, or handle the work?

Teams often solve this with:

  • Names embedded in prompts
  • Static email addresses
  • Environment variables
  • YAML routing tables
  • Slack-channel conventions
  • Custom directory lookups
  • Workflow-specific approver rules

Those approaches duplicate organizational logic across every agent. Roster centralizes the participant model and makes it callable through MCP.

How the integration works

MCP is an open standard for connecting AI applications to external tools, data, and workflows. Roster uses MCP as its primary agent-native integration surface.

AI agent
   ↓ MCP tool call
Roster /mcp
   ↓
Project and participant selection
   ↓
Directory membership and metadata
   ↓
Active delegations
   ↓
Resolved participant and users
   ↓
Agent continues its workflow

Use resolve for open-ended questions, exact tools for exact checks

resolve
Accepts a natural-language query, an optional project ID, and an optional test or live mode. Roster extracts relevant project references, labels, roles, and constraints, then resolves associated users, active group members, and active delegations.
Who should approve the Atlas vendor renewal?
Projects and participants
List projects available to the acting identity. List or inspect participants in a project.
List participants in the Atlas project.
Membership checks
Deterministically verify whether a user is a member of a participant.
Check whether this user is a member of the Release Approver participant.
Delegations
List active delegations that affect a participant right now.
List active delegations for this participant.
Directory records
Look up materialized directory records associated with participants.
Resolve request history
Retrieve recent Resolve requests for audit, context, or observability.

Authentication modes

Roster MCP supports both OAuth (recommended for interactive human users connecting through AI clients) and Roster API keys (recommended for service accounts, autonomous agents, CI pipelines, and non-interactive automation).

Roster API keys use the rst_ prefix, belong to a Roster identity, carry explicit scopes, can be inspected and revoked, and remain subject to the owner's Roster permissions. The default Roster MCP authentication mode can allow both OAuth and API-key access.

Resolve-only service-account scope — a narrowly scoped agent may only need:

mcp:resolve
mcp:resolve-requests:read
mcp:projects:read
mcp:participants:read

Do not grant project, participant, label, delegation, or administrative write tools unless the agent is deliberately responsible for maintaining those resources.

Agent-framework pattern

Roster can be used with any framework or runtime capable of connecting to a compatible remote MCP server. Use a system or skill instruction such as:

Use Roster whenever the workflow needs to determine who should approve,
own, review, handle, receive, or be contacted for a task.

Call `resolve` for open-ended organizational-routing questions.

Use exact Roster tools for project, participant, membership, label,
delegation, and Resolve-history checks.

Treat Roster as the source of truth for current participant assignments
and delegations.

Never invent an owner or approver when Roster returns no authorized result.

Tool descriptions and agent instructions are important because connecting an MCP server does not guarantee that an agent will select its tools for every relevant prompt.

Example agent workflow

An autonomous procurement agent has extracted a vendor request, classified the purchase, checked the contract amount, and identified the relevant business project. Before creating an approval request, it calls Roster.

Who should approve this European software purchase for Atlas Operations at this amount?
  1. The agent calls resolve with the classified business context:
    resolve(
      query:
      "Who should approve this European software purchase
       for Atlas Operations at this amount?"
    )
  2. Roster evaluates the organizational context and returns the appropriate participant and active users.
  3. The agent sends the task through the organization's existing workflow or messaging system.
  4. After the decision, the agent continues its workflow.

Roster identifies the participant. Another system delivers, records, and enforces the approval.

MCP versus the REST API

Use MCP when the caller is an AI agent, natural-language tool selection is useful, the runtime already supports MCP, the agent needs multiple Roster tool families, or you want standardized tool discovery.

Use REST when the caller is a conventional application, the workflow is fully deterministic, you want direct endpoint-level control, your platform does not support MCP, or you are building an SDK or backend service.

Both surfaces use Roster's governed project, participant, directory, membership, and delegation model.

Security design

Identity attribution
Every agent or service account should act through its own Roster identity.
Least privilege
Grant only the scopes required for the workflow.
Project boundaries
MCP scopes do not bypass project and resource-level authorization.
Controlled context
Send only the business context needed for the routing decision.
Explicit failure handling
Treat no-result, ambiguity, authorization failure, and provider failure as workflow states rather than opportunities for the agent to guess.
Auditable requests
Use Resolve history and available audit capabilities to inspect how organizational decisions were made.
Human handoff separation
Roster identifies the participant. Another system should deliver, record, and enforce the approval or task.

Frequently asked questions

No. Roster is a participant-resolution platform and MCP server that agents and frameworks can call.

Give every agent the same organizational “who” primitive

Use one governed participant-resolution surface across agent clients, internal agents, automation platforms, and CI workflows.

Connect an MCP clientExplore the MCP toolsExplore all integrations