mcp-data-gateway

Hand an AI client to one person and MCP is easy. Hand it to a team — analysts, people-ops, leads, each entitled to different systems — and the server needs an answer it can defend: who is calling, and what may they touch? This repo is that answer, cut down to its load-bearing parts.

One allows() decision drives everything. It filters what each identity sees and refuses what they try anyway, so the tool list is never the only thing standing between a caller and a system. Grants fail closed: a missing or corrupt store denies everyone except a break-glass superuser who lives in code, precisely so no store outage can lock out the person who fixes stores. Every call — and every denial — lands as a size-capped audit line.

Opinions baked in:

• Grants name tools one by one. Globs get rejected at load time, because the write tool someone ships next quarter must not inherit its way into a read-only identity.

• A tool that joins two systems declares both, and the caller needs both. The reads are what leak, not the names.

• A snapshot test pins the tool surface, and a second one holds the write surface at exactly one tool.

• The SQL gate gives fast, specific refusals — but the real fence is the database identity's permissions. You can talk past a text scan. You cannot talk past IAM.

Quickstart

python3 -m unittest discover -s tests   # no dependencies
python3 -m gateway.demo                 # three identities, one denial on the record
pip install mcp && GATEWAY_IDENTITY=analyst@example.com python3 -m gateway.server

Related MCP server: Agentrim MCP

Going deeper

docs/design.md walks the pieces with captured output. docs/rules.md gives every rule its reason and its test. docs/lineage.md holds the honest limits and provenance.

Distilled August 2026 from production gateways that earned each rule the hard way. The commit log starts at the distillation.