Setoku MCP Server – Data Knowledge for AI Agents

TL;DR

TL;DR: Setoku is an open-source MCP server that sits between your AI and your company data, feeding it metric definitions and operational context before every query — so the AI computes numbers the way your business actually does, not by guessing from column names.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: setoku.com
• Source repository: github.com/Hedgy-Labs/setoku (7 stars, Apache-2.0, verified via GitHub API)
• License: Apache-2.0 (verified via GitHub API license.spdx_id)
• HN launch thread: not available (3 points on HN, no dedicated launch thread found)
• Source last checked: 2026-07-24 (commit main pushed 2026-07-23)

What Is Setoku?

Every company has knowledge that lives in people’s heads: which metric is the real one, why “paying customer” is trickier than it looks, why a naive COUNT(*) returns the wrong number. AI agents never had that — they guess and get it confidently wrong.

Setoku is the curated institutional memory layer underneath your AI. It holds metric definitions, data gotchas, and schema documentation, then feeds them to your AI right before it answers a question about your data. The agent looks up the context first, then queries, so it computes things the way your company actually does.

The MCP server ships tools, not models. No AI inference runs on the Setoku server itself — the thinking happens in the AI you already pay for (Claude, Codex, or any MCP client). That means no added inference cost: a whole deployment is one small VPS.

From the README:

Setoku = set (math) × oku (奥, innermost): the innermost layer underneath your AI.

Three Tool Categories

Setoku exposes three kinds of MCP tools:

Context Tools

• find_context — retrieve canonical metric definitions and data documentation
• get_metric — fetch a specific metric with its definition and known gotchas
• report_correction — propose a change to what Setoku knows (a human approves on the admin page, outside the agent loop)

Query Tools

• get_schema — inspect the database schema
• run_query — execute a read-only query against the analytics engine, with a row cap, statement timeout, and append-only audit log

App Tools

• publish_app — turn an answer into a small self-contained web app on live data
• update_app — update an existing published app

How Deployment Works

Setoku runs on a single VPS. The README covers the full self-hosted setup:

1. Deploy the server (one small VPS)
2. Connect data sources (Postgres, MySQL, BigQuery, or a data lake)
3. Define metrics and gotchas via the web admin console
4. Share the MCP server URL with your team — each person gets per-source credentials, revocable from the web console

The gateway never holds credentials to your production database. A mirror job does (read-only), and both read-only and per-source access are enforced by the database engine, not by parsing SQL in Setoku’s code.

The Apps Feature

Once your AI understands your data, it can build little things on top of it. Ask for a chart, a triage list, or a poll — the agent writes a self-contained app and hands you a URL. Non-technical teammates can describe what they want in plain language and edit it the same way.

Apps run in a locked-down, no-network sandbox backed by live data through the same governed query path. A prompt-injected app can mess up its own notes but cannot write to your sources.

Security Model

• Read-only by design. The agent only runs audited, read-only queries against a mirror of your database.
• Per-person credentials. Connector URLs are per-person database connection strings, revocable from the web console.
• Human approval gate. The agent can propose changes to metric definitions, but a human accepts them on the admin page — outside the agent loop.
• No production credentials on the gateway. The gateway never holds a credential to your production database; a mirror job holds the read-only credential instead.

Full details: SECURITY.md in the repo.

Try the Live Demo

Setoku hosts a public demo wired to a synthetic dataset for a fictional pro sports club, the Bonita Bulldogs — covering ticketing, CRM, sponsorship, merchandise, staffing, payroll, and more.

To connect the demo to Claude (or any MCP client):

1. Open Settings → Connectors → Add custom connector
2. Paste this server URL:

   https://demo.setoku.com/mcp/fdb6bb54d746ba8e00d698ff2183228b682b8272bfef78e0

3. Ask a question in plain language, for example:
   • “How many unique fans do we have?” → 71,204 (deduped by normalized email, not the raw 92,118 a naive COUNT(*) returns)
   • “What was our ticket revenue this season?” → $46.8M (cents reconciled, refunds/exchanges/comps excluded)
   • “What’s our season-ticket renewal rate?” → spans three seasons of ticketing history

The demo data includes live published dashboards — Sponsorship pricing table and Fan lifetime value.

FAQ

Q: Does Setoku run AI models on its server? A: No. Setoku is a tools and context server only. No AI inference runs on the Setoku server; all reasoning happens in the AI client you already use. This keeps deployment cheap — a single small VPS, plus your existing AI seats.

Q: How does it differ from a direct database MCP server? A: A plain database MCP server exposes raw tables and columns. Setoku adds a curated context layer: metric definitions, known gotchas, entity documentation. Before running a query, the AI receives the business logic it needs to compute numbers correctly instead of guessing from column names.

Q: Can I self-host it? A: Yes. The server is open source (Apache-2.0) and deploys to a single VPS. Data source connectors support Postgres, MySQL, BigQuery, and data lakes.

Q: Can the AI rewrite the metric definitions? A: No. The agent can propose corrections via report_correction, but a human must approve changes on the admin page. The agent cannot modify what Setoku knows without human oversight.

Q: What license is it released under? A: Apache-2.0, verified via the GitHub repository license.spdx_id field.

Conclusion

Setoku solves the “confidently wrong AI” problem for company data. By feeding metric definitions, schema context, and operational gotchas to your AI before every query, it turns naive COUNT(*) queries into business-accurate answers — without adding AI inference costs, since no model runs on the server itself.

The MCP protocol means it works with any MCP client: Claude, Codex, or anything else. If your team asks AI questions about company data and has ever gotten a confidently incorrect answer, Setoku is worth a look.

Project: setoku.com · GitHub · Live Demo