Runs locally today
The server uses a local stdio transport for an MCP client. There is no hosted/remote endpoint yet — that's on the roadmap.
Feature · MCP
Let AI agents work with your data — safely.
KIN Create ships a read-only Model Context Protocol (MCP) server so agents like Claude can read the right data on a user's behalf — under the same tenant isolation and permission rules as the app, and with nothing ever written back.
For developers & operators
Read-only by design
Every tool is side-effect-free. The MCP surface can never write, mutate or delete — only read.
Tenant-isolated
Access is scoped to the caller's workspace; cross-tenant reads return a non-disclosing response.
Same permission matrix
Per-tool authorization reuses the app's role/permission rules — no bypass of the app layer.
Explicit, revocable access
Served only for a principal with an active access grant; revoking a grant fails closed immediately.
Rate-limited & audited
Per-user, per-tenant, per-tool rate limits, with immutable audit evidence that carries no payload.
Prompt-injection resistant
Retrieved records and arguments are treated as inert data — stored text is never interpreted as instructions.
Step by step
How to connect an agent
- 1
Enable the server
Turn on the feature-gated MCP server for your environment (off by default).
- 2
Grant access
An explicit, revocable access grant is issued for the principal and tenant.
- 3
Connect a local client
Point an MCP-capable client (e.g. Claude) at the local stdio server; identity travels with the session, not tool arguments.
- 4
Read safely
The agent calls read-only tools; output is bounded, minimized and redacted, and every call is audited.
1. What the agent can read
The server exposes a small, deliberate set of read-only tools. Each returns a bounded, minimized, redacted view — never a raw dump.
- kincreate.campaigns.summary — the caller's own campaigns (id, name, status, budget)
- kincreate.confidence.read — a creator's explainable audience-confidence score with factors and limitations
- kincreate.creator.profile — a creator's public, discoverable marketplace profile
2. Security enforced at the boundary
Every control lives at the trusted server boundary, never in a tool: the feature gate, identity verification, access-grant check, per-tool authorization, rate limiting, output shaping and audit. One tools/call invokes exactly one named tool.
Because the MCP server reuses the platform's own tenant isolation and permission matrix, an agent can never see more than the user it acts for is allowed to see.
3. Built for the AI-native era
MCP is the emerging standard for connecting AI assistants to real systems. Shipping a governed, read-only MCP server means KIN Create works with agent workflows out of the box — without handing an agent write access or exposing data across tenants.
Frequently asked questions
- Can an AI agent change my data through MCP?
- No. The MCP server is strictly read-only and side-effect-free — there are no write, update or delete tools.
- Does it respect my permissions?
- Yes. Per-tool authorization reuses the same role/permission matrix as the app, and access is tenant-scoped with a non-disclosing response for anything out of scope.
- How is access controlled?
- The server is feature-gated and off by default. Access is served only for a principal holding an active, revocable grant; revoking it fails closed.
- Is it safe against prompt injection?
- Retrieved records and tool arguments are treated as inert data — the server never interprets stored text as instructions, and each call runs exactly one named tool.
- Which agents can connect?
- Any MCP-capable client, including Claude, over the local stdio transport. A hosted/remote transport is on the roadmap; see the developer docs.
Ready to try let ai agents work with your data — safely.
Explore more: Developer docs · AI assistance · Analytics & tracking