AI agents
Ask your AI assistant about HR, and get answers from HumanR
HumanRspeaks the Model Context Protocol, the open standard AI assistants use to reach the systems you work in. Connect the assistant you already use and ask it to catch you up, clear your approvals, book your leave or total last year’s deductions by company. It acts as you and sees only what you can see. Every change it makes waits for your yes.
Which assistants can connect today
Each person connects as themselves: by signing in to HumanR from the assistant, or with an agent token they issue inside HumanRand paste in with your site’s address.
Claude Code
Works today
One command with your site address and a personal token.
Cursor
Works today
An entry in ~/.cursor/mcp.json with the address and the token as a header.
VS Code (GitHub Copilot agent mode)
Works today
A .vscode/mcp.json that asks for the token when the server starts, so it never sits in a file.
claude.ai, Claude Desktop connectors and ChatGPT
Works today
Add a custom connector with your site address. You sign in to HumanR, see which app is asking, and tick what it may do. No token to copy.
Microsoft Copilot Studio
Works today
Add a Model Context Protocol tool. Either sign each person in with OAuth, using a client your admin registers in HumanR, or use one API key in the Authorization header.
Microsoft 365 Copilot (Chat, Teams)
Works today
Through a Copilot Studio agent your admin builds and publishes to the organisation. With OAuth sign-in each person acts as themselves; with a single key everyone acts as one identity, so keep that for a pilot.
Things people ask it
Plain questions, answered from your live records. What each person gets depends on what they may see in HumanR: an employee hears about their own leave and payslips, and a payroll lead can ask about the whole company.
Catching up
- “Catch me up on HumanR since Monday.”
- “What is waiting on me?”
- “Anything coming up next week I should know about?”
Self-service
- “How much annual leave do I have left?”
- “Book me leave from the 12th to the 18th of November.”
- “Show me my last three payslips.”
Approvals
- “What is in my approval inbox, oldest first?”
- “Approve Aisha's overtime claim for Saturday.”
HR and payroll questions
- “Who is away today, and who should already be back?”
- “How much did our two companies deduct for staff loans and advances last year, by currency?”
- “Top ten people by food allowance paid this year.”
- “Record sick leave for staff number FS-870 on 30 September.”
What stays true when an agent is connected
- The agent is the person, never more
- Every call runs as the person who connected it, with their live permissions. Lock their account, remove a role or change a permission, and the agent loses it within a minute. The same masking that hides pay and personal fields in the app hides them from the agent.
- A token can only narrow
- A personal token carries capabilities: read, act on my own requests, approve, see my own payslips, and record for others. Each one is ticked by the person when they issue it. A token can also be limited to one company. None of them adds a permission the person does not already hold.
- Nothing changes without a yes
- Every change the agent makes is shown back first, with dates in words and the employee's name on record, and needs a second confirming call. A retried confirmation never writes twice. Clients that support it get a native confirmation prompt instead.
- Every call is on the audit trail
- Each tool call is logged with the person, the agent, what was asked and the outcome, and never the token. Admins see every connected agent, can revoke any of them, and have a switch that stops all agent access at once.
- Employees only when you say so
- Staff can connect an agent once the module is on. Employees who use only the self-service portal cannot connect until an admin allows it. That setting is off by default, and turning it off again disconnects their agents.
- Text from records is treated as data
- Free text from records, such as remarks, reasons and notes, reaches the agent fenced and labelled as data, so an instruction typed into a leave reason is not obeyed as an instruction.
Connecting takes a minute
Issue a token under Connected apps (or My assistants on the employee portal), choose what it may do, and paste it in. The page shows the token once, along with these settings already filled in with your address. The address is your HumanR address followed by /mcp, on our cloud, your own domain or your own servers.
claude mcp add --transport http humanr https://hr.yourcompany.com/mcp \
--header "Authorization: Bearer hr_pat_…"{
"inputs": [
{ "type": "promptString", "id": "humanr-token",
"description": "HumanR agent token", "password": true }
],
"servers": {
"humanr": {
"type": "http",
"url": "https://hr.yourcompany.com/mcp",
"headers": { "Authorization": "Bearer ${input:humanr-token}" }
}
}
}Calls per token
120 a minute
Calls per site
1,000 a minute across every agent
Token lifetime
90 days by default, up to a year
Tokens per person
10
Rows per answer
Up to 100, with a cursor for the next page
Default limits. They can be changed for a site that needs more.
Questions about AI agents
Does HumanR send our data to an AI company?
No. HumanR answers the questions your people's agents ask, and only with what each person may already see. The answer then goes to whichever AI provider that person's tool uses, under the terms your organisation has with that provider. You remain the controller of your HR data. Decide which tools your staff may use, the same way you would for email or spreadsheets.
Which AI model does it use?
None of its own. HumanR is the source of answers, not the model. Claude, GPT or whatever your client runs does the reading and the writing; HumanR supplies the records, the totals and the confirmations.
Can an agent run payroll or change salaries?
No. The changes an agent can make are deliberately few: submit and cancel your own leave, overtime and expense requests, approve or reject what is in your inbox, mark notifications read, and, for HR with the right permission, record leave, sick leave, overtime and expenses for an employee. Payroll runs, salary changes and employee records stay in the app.
How does it get money questions right?
It answers through the same report builder your finance team uses, over finalised payroll runs. It never adds two currencies into one number without converting them, and when it converts it says at which rate. It names the period it used, so 'last year' never quietly means the last twelve months. Pay terms are looked up before they are counted: 'loans' covers every loan and advance item, and an allowance paid through a bank channel is found as a channel, not as a pay item.
We are on our own domain, or self-hosted. Does it still work?
Yes. The address is your HumanR address followed by /mcp. The token pages inside HumanR show it filled in, with ready-to-paste settings for each client.
How do we turn it on?
Ask us to switch AI agents on for your site. Staff then connect by signing in from claude.ai or ChatGPT, or issue their own tokens under Connected apps. If you want employees on the self-service portal to connect, an admin allows that on the AI agents admin page.
See it answer from a real instance
Book a demo and we will switch AI agents on in your demo company, so you can connect your own assistant and ask it anything.
No credit card. No sales call required. A real login, emailed to you.