Developers
MCP server for AI assistants
Let Claude and other MCP clients answer questions about your floor and record interactions, with the asking person’s own access. Tools, resources, prompts, client configuration and the security model.
What it is
The Model Context Protocol (MCP) is an open standard that lets AI assistants call tools in other systems. TINIUN runs a remote MCP server at https://tiniun.com/api/mcp (Streamable HTTP, JSON responses, no sessions). Ask your assistant things like “Who has been on break the longest?”, “Which team is short-staffed tomorrow morning?” or “How many tickets did Maria reply to this week?”
It uses the same scopes, permissions and team scope as the REST API: tools, resources and prompts you haven’t been granted aren’t even listed.
Connecting a client
Use OAuth. MCP clients register themselves, open the TINIUN consent screen, and receive a token that acts as you. You don’t create or paste anything.
Add the server
Give your client the URLhttps://tiniun.com/api/mcp.Approve
Your client opens TINIUN in the browser. Sign in, choose the organization and approve the scopes.Ask
The tools appear in the client.
claude mcp add --transport http tiniun https://tiniun.com/api/mcp
# then run /mcp inside Claude Code and choose “Authenticate”claude mcp add --transport http tiniun https://tiniun.com/api/mcp \
--header "Authorization: Bearer spk_your_key_here"Settings → Connectors → Add custom connector
Name: TINIUN
URL: https://tiniun.com/api/mcp
(then choose Connect and approve in the browser){
"mcpServers": {
"tiniun": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://tiniun.com/api/mcp"]
}
}
}{ "mcpServers": { "tiniun": { "url": "https://tiniun.com/api/mcp" } } }curl https://tiniun.com/api/mcp \
-H "Authorization: Bearer spk_your_key_here" \
-H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'Menu names differ between clients and change over time; follow your client’s documentation for adding a remote MCP server with OAuth. For headless automation that can’t open a browser, send an API key instead: header Authorization: Bearer spk_…. Prefer OAuth wherever a person can approve.
Tools
| Tool | Needs scope | Cost | What it does |
|---|---|---|---|
| whoami | none | 1 | The organization, credential type, scopes and rate limits of this connection. |
| list_people | people:read | 1 | People with email, title, team and status; search by name or email. |
| list_teams | people:read | 1 | Teams and their minimum staffing. |
| live_status | status:read | 1 | What every visible agent is doing right now, and since when. |
| list_schedules | schedules:read | 1 | Published schedules by team and period. |
| list_shifts | schedules:read | 2 | Shifts with breaks and meals for a period of up to 31 days. |
| list_exceptions | exceptions:read | 1 | Late logins, break overages and other deviations. |
| adherence_daily | adherence:read | 3 | Daily scheduled, actual, productive and adherent time per agent. |
| list_leave_requests | leave:read | 1 | Time-off requests and their status. |
| list_holidays | holidays:read | 1 | The organization’s holidays for a period. |
| forecast_staffing | forecast:read | 5 | Forecast volume and agents required against scheduled, by queue and interval, for one day. |
| advisor_summary | advisor:read | 10 | Coverage gaps and burnout risk counts, without individual-level data. |
| list_interactions | interactions:read | 2 | Tickets replied to, conversations handled and calls answered. |
| record_interactions | interactions:write | 5 | Record interactions handled by agents (idempotent). |
Every tool has a human-readable title and the annotations clients use to decide how much to ask before running it: readOnlyHint (true for everything except record_interactions), destructiveHint: false, idempotentHint: true and openWorldHint: false. Results come back both as text and as structuredContent that matches the tool’s outputSchema. Cost is the number of calls a tool use counts against your per-minute budget; when it is used up the tool returns a readable “rate limit reached, try again in N seconds” error. Results over about 60,000 characters are cut at a row boundary and stay valid JSON, with a nextCursor to continue.
Resources
Resources are data an assistant can read without a tool call. They are listed with resources/list and fetched with resources/read, only when the credential holds the scope.
| URI | Needs scope | Contents |
|---|---|---|
| tiniun://status/live | status:read | What every visible agent is doing right now. |
| tiniun://schedules/today | schedules:read | Shifts that overlap today in the organization’s time zone. |
Prompts
Prompts are ready-made instructions that a client can offer as shortcuts (prompts/list, prompts/get). A prompt is listed only when the credential holds every scope it needs. Arguments are validated before they go into the text.
| Prompt | Needs scopes | What it asks the assistant to do |
|---|---|---|
| daily_briefing | status:read, schedules:read, exceptions:read | A short morning briefing: who is on, what is open, what needs attention. |
| burnout_review | advisor:read | Review burnout risk across teams and suggest what to change. |
| coverage_check | advisor:read | Find upcoming coverage gaps and propose fixes. |
Protocol and limits
- Revisions: 2026-07-28, 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05. The current revision,
2026-07-28, is stateless: every request names its version in_metaand in theMCP-Protocol-Version,Mcp-MethodandMcp-Nameheaders, andserver/discoverreturns the server’s versions, capabilities and instructions. Earlier revisions use theinitializehandshake. Both work on the same URL; clients pick automatically. - Capabilities: tools, resources and prompts, with no list-change notifications (the lists depend only on your scopes) and no subscriptions.
GETandDELETEon the endpoint answer 405. - Size: requests up to 100 KB; batches of up to 20 messages on the earlier revisions.
- Rate limits: the same per-credential budget as the REST API, with each tool spending its cost. Rate limit headers are on every response.
- Errors: problems with a tool call come back as a tool result with
isError: trueand a readable message; protocol problems use JSON-RPC error codes (-32602bad parameters,-32601unknown method,-32022unsupported protocol version,-32020header mismatch).
Security model
- Connecting an AI assistant sends the data it asks for (names, schedules, statuses, figures) to that assistant’s provider under your agreement with them. Approve only assistants your organization trusts, and only the scopes they need.
record_interactionsis the only tool that writes. Everything else is read-only.advisor_summaryreturns team-level counts and gaps, never names.- Treat data returned by tools as data, not instructions. TINIUN’s own “Ask TINIUN” assistant does the same.
- Audit: every tool call and resource read is written to the audit log with the credential, the tool, how long it took and whether it worked. Only the names of the arguments are recorded, never their values, because values can be personal data.
- Admins see every connected assistant, who approved it and when it was last used, and can disconnect it instantly under Settings → Integrations → OAuth applications. Usage per key and application is under Settings → Integrations → API usage.
- Browsers: pages on other sites cannot drive the endpoint. A request carrying an
Originheader that isn’t TINIUN’s own (or one of the deployment’s configured browser origins) is refused with 403 before authentication, which defeats DNS-rebinding attacks. Native clients send noOriginand are unaffected. - An unauthenticated call gets
401with aWWW-Authenticateheader whoseresource_metadatalink (RFC 9728) tells the client where to find the authorization server, which is how MCP clients start the OAuth flow by themselves.
See it with your own scenarios
A 30-minute walkthrough with our team, tailored to your operation.
Book a demo
