← The ledger
Receipt 002Agent infrastructure

57 governed tools, 654 tests, one live MCP server

The falsifiable questionCan one MCP server carry organizational memory, bounded authority, and receipts to every major AI client — and stay up in production?

ARunnable artifactearned
BMeaningful baselinewithheld
MMeasured resultearned
RReproducible methodwithheld
FDocumented failureearned
EExternal validationearned

Baseline and reproducible withheld: the cold-start baseline is argued, not measured, and the server is a product, not a rerunnable harness. The benchmark receipt covers that ground.

Baseline

Without a continuity layer, each AI client starts cold: the goal, prior decisions, quality bar, and permissions live in the human's head and get re-typed into every session. The human is the API between their own tools.

Method

A Cloudflare Worker MCP server (orgx-mcp) exposing 57 tools across memory, planning, delegation, decisions, receipts, and lifecycle — with OAuth, session isolation, and durable state. Distribution manifests shipped for Smithery, Glama, and the ChatGPT apps directory. Test suite: 654 automated cases across ~90 spec files.

Measured result

Live at mcp.useorgx.com with a public .well-known/mcp.json manifest. 57 published tools, 654 passing tests, and a public status page showing 98.271% uptime since April 2026 — the incidents are on the record, not hidden. Repos under github.com/useorgx have external stargazers.

Where it broke

98.271% is not 99.9%: real outages sit on the public status page. And GitHub star counts are near zero — distribution manifests shipped, but registry-driven adoption is still an open bet, not a result.

Next test

Publish per-tool success-rate and latency numbers from production traffic.