Ask your society
a question.
An assistant in the resident app and the admin portal that answers from your live community data — in English, Hindi or Telugu. Behind it, a Model Context Protocol server with 70 tools, so your community is addressable by an AI client at all.
The same assistant, wherever you are.
Not a chatbot bolted onto a help centre. It reads the same data the app reads, with the same permissions you already have.
70 tools. 11 domains.
MCP is the open standard for giving an AI client structured access to a system. AKCS ships an MCP server, so a community running on AKCS can be read and acted on by an AI client directly — with the same role scoping and the same approval step a person gets. As far as we know, no other society platform in India exposes one.
55 read, 15 write. Adding a tool is one decorated function, so the catalogue grows with the product rather than lagging behind it. Developer reference
What stops it doing something stupid.
An assistant with access to a society’s gate, ledger and resident records is only worth having if you can say precisely what constrains it. Here is the whole list — ask any other vendor for theirs.
Nothing is written without a human
Every one of the 15 tools that change something is two-phase. The model proposes; you get a card showing the exact arguments; the tool runs only after you accept. The approval token is signed, bound to you and that conversation, and expires in 15 minutes.
It cannot call what it cannot see
The tool catalogue is scoped by role. A resident’s session does not contain the admin tools at all — they are absent from the list, not merely refused.
Prompt injection cannot cause a write
Complaint text, visitor names and announcements reach the model as tool output, which the system prompt designates as data and never as instructions. Because writes need approval regardless, injected text has no path to a mutation. A live test plants "IGNORE ALL PREVIOUS INSTRUCTIONS" inside a complaint and asserts nothing is written.
It cannot invent an answer
The opening turn of every exchange is forced to call a tool. The assistant answers from your data or it does not answer — it has no way to reach for a plausible-sounding number.
Every call is on the record
Each invocation records the tool, the arguments, the result, whether a human approved it, and how long it took. An AI acting on a society’s money and gate needs an audit trail, so it has one.
You can switch the writing off
One setting strips all 15 mutating tools from the catalogue, leaving the assistant strictly read-only. Another disables it entirely. Committees that want answers without actions can have exactly that.
“Approve the visitor
with AKCS.”
Said out loud, phone in your pocket, guard waiting at the gate. iOS 16 and later.
In Mumbai, on your data, on the record.
Indian region, by default.
The assistant runs on AWS Bedrock in ap-south-1 — Mumbai. Your residents’ questions and the data behind the answers stay in the region.
- Answers grounded in a tool call, never invented
- Per-user and per-community daily limits
- Every tool call recorded with its arguments and result
- Read-only mode available for committees that want it
