MCP server/settings/mcp

Let an AI assistant query your analytics

Asking "which channel converted best last month" is faster than building the report. Falorb ships an MCP server so any assistant that speaks the protocol can answer it directly from your data, running on your machine, under a key you issued, with the destructive operations deliberately absent.

Install Falorb See the whole product

The MCP settings page: connection configuration for an assistant beside a list of issued API keys with their scopes. — Connect an assistant · API keys

Connect it in one block of configuration

stdio transport for local clients (Claude Desktop, Claude Code, anything running on the same machine), or streamable HTTP with bearer-key auth for remote and hosted clients. The dashboard shows the exact configuration to paste.

A connect-an-assistant panel showing the MCP client configuration block to copy.

Tools that cover the product, not a demo subset

Discovery, analytics, funnels, people, live and ops: 25 tools in total, each verified against a real MCP client. Output is shaped for a language model: markdown tables, numbers already formatted, times expressed relatively.

Discovery: list projects, event names, property keys, describe filters
Analytics: overview, stats, trend, breakdown, retention, stickiness, drop-off, user flows
Funnels: run a funnel, get its drop-offs
People: list, search, full profile, cross-project, sessions
Live and ops: live visitors, event stream, platform health, alerts, install snippet

Isolation that was tested, not assumed

A second organisation's key sees only its own projects and cannot name another tenant's. A read-only key is refused a write tool. Each request gets its own server instance so scope cannot leak between concurrent callers. All three are covered by the smoke suite, which runs in CI against real services.

An API keys panel listing issued keys by prefix, with scopes, creation date and revocation controls.

What it deliberately cannot do

There is no project deletion and no person erasure over MCP. Irreversible actions should not be reachable by an assistant acting on a misread instruction, and erasure in particular needs a human to confirm whose data is being destroyed. Creating an alert is the one write tool, and it requires a key scoped for writes.

Specifics

How it works, precisely

Transports: stdio, and streamable HTTP with bearer API-key auth
Tools: 25, plus 2 resources and 3 prompts
Prompts: weekly_review · conversion_audit · lead_research
Ranges accepted: 7d, 24h, today, mtd, or 2026-08-01..2026-08-16
Not available: Project deletion, person erasure. Excluded by design

Questions

MCP server: the usual questions

Does my analytics data get sent to an AI provider?

Only what the assistant asks for and only if you connect one. The server runs on your infrastructure; the assistant calls it like any other tool, and the results enter that conversation. If you connect nothing, nothing leaves.

Does it support OAuth?

Not yet. Authentication is bearer API keys, which is what most connectors accept. OAuth for MCP is a known gap rather than a design decision.

Next, most people read

AI & crawlersEvery AI agent that fetched your pages, split by whether it was answering a question or building a corpus, plus the readers they sent back.Teams & sharingOwner, admin, member and viewer roles re-derived on every mutation, plus invitations and public read-only links.Cross-project insightsA metric × dimension × chart builder across all properties, plus the people who used more than one of them.

Run it yourself this afternoon

Clone the repository, bring up the stack with Docker Compose, paste one script tag. The data never leaves your machines.

Quickstart View on GitHub