Agents
Agents run it. People decide.
Your agents order checks and explain delays. They can’t sign consent, read reports or reject a provider.
Request
create_check(
provider="maya@example.com",
package="care_standard"
)Response200 OK
Consent link sent to Maya.
Only Maya can sign it.An example agent call. Names are made up.
What an agent can and can’t do.
Three things it can do. Three places it stops.
An agent can start a check, explain a delay in plain words and tell you which checks are still open. That covers most of what your ops team gets asked.
It stops at consent, at the report and at adverse action. Each refusal says what happened and who can do it instead.
“Start a check for Maya”
Agent can do itConsent link sent to Maya.
“Tell Maya why it’s slow”
Agent can do it“Waiting on Hamilton County. About one more day.”
“Which checks are still open?”
Agent can do it“Three open. Two waiting on consent, one on Hamilton County.”
“Sign the consent for Maya”
Only the providerOnly Maya can sign.
“Reject a provider”
Only a person on your teamA person on your team approves this.
“Read me the full report”
Only keys with reports:readStatus-only key. Your ops team reads reports.
Agents get a status-only key.
The key decides what an agent can reach, not the prompt.
Give each agent its own key with checks:write and checks:read. It can start checks and read status, by search and by court. It can’t open a report: that needs reports:read, which you keep for your ops team. The report itself is signed by the analyst who confirmed each record at the court.
Each agent’s key is separate, so you can rotate or revoke one without touching your onboarding service.
Request
get_report(check="chk_8Rq2")Response403 Forbidden
This key is status-only. Use a key with
reports:read, or open the report page.A person approves adverse action.
No key, agent or prompt can reject a provider.
Roster prepares the pre-adverse and adverse letters with the waiting-period dates filled in, and sends a webhook when the period ends. The adverse letter stays held until a person on your team approves it. Roster enforces that gate on every key.
- PreparedPre-adverse letterYour team sends it
- Dates filled inWaiting periodWebhook when it ends
- Your callA person approvesNo agent can
- After approvalAdverse letterYour team sends it
Works with Claude, ChatGPT, Gemini and Copilot.
Any client that speaks MCP. Each one knows where to stop.
The limits live in Roster, not in the client. Whichever assistant your team uses, the same tools come back with the same refusals.
Tool descriptions say what each tool does and what it can’t, so the agent can tell your team before it tries.
- ClaudeDesktop and Claude Code
- ChatGPTConnectors
- GeminiGemini CLI
- CopilotAgent mode in your editor
Connect over MCP.
Add the roster server with a status-only sandbox key.
Add the server to your client’s MCP config with a sandbox key. MCP access opens with the sandbox. Start in sandbox with test providers, then swap in a live status-only key after your credentialing call.
Building your own agent? The same tools are on the API. See the tool list.
MCP config
{
"mcpServers": {
"roster": {
"url": "https://mcp.rosterchecks.com",
"headers": {
"Authorization": "Bearer sk_test_status_…"
}
}
}
}What Roster handles
Sandbox keys right away, no sales call. Live keys after a short credentialing call.
- One call starts a check. Hosted consent and live ID in your brand; a webhook when it’s done. Sandbox keys without a sales call.
- Status by API, webhook and MCP. explain_status says which court it’s waiting on and when to expect it.
- A named analyst matches every possible record to the provider and confirms it at the court before it’s reported. A signed report page and PDF your team can open.
- Every notice and letter prepared, with the waiting-period dates filled in, and a human approval gate before adverse action.
- You send the notices and make the final call. Contractors covered the same way as employees.
Reviewed November 5, 2026 by Ben Drellishak, Owner.