Looph

The prototype ran on Lovable's own hosting, compiled for Cloudflare Workers, with no container, no health check and no second copy to try a change on before customers saw it.

A platform of its own

Looph now builds into a Node server in Docker, with a health check that proves the database answers, not only that the process is up. Production runs on DigitalOcean App Platform from its own release branch, and staging is a full second copy with its own database and test-mode payments, sign-in and email. Database migrations never ride along with a deploy: they are applied by a separate, deliberate step.

The browser talks to the database directly with the signed-in person's token, so a permission checked only on screen is no permission at all, and one team's feedback must never reach another's.

Every workspace walled off in the database

Looph's database checks every read and write itself, row by row. Row-level security is on for all 172 public tables, and 1,180 policies decide what a browser, an API key, an AI agent or an anonymous visitor may read or write. One database function answers every permission question for every caller, so there is no second copy to drift. 110 pgTAP suites, tests that run inside a real database, cover those rules, and three privileged functions that any signed-in user could call were found and closed, checked on production before and after.

Teams want the board on their own domain. It shares no cookie with Looph, so visitors would sign in again or the sign-in provider would charge per domain, and a hostname typed into a form must never go live before its owner proves it.

Their own domain, still signed in

Every workspace gets its own subdomain, and a customer's domain is served through Looph's own edge, with certificates issued automatically on the first request. Only a domain that has passed its DNS check resolves at all. On it, people stay signed in through Looph's own first-party session: a single-use, sixty-second handoff code, stored only as a hash, and a cookie that never outlives the sign-in behind it. In production since 17 September 2026.

The whole product is one promise: whoever asked hears when it ships. When we measured the email pipeline, 23 notifications pointed at templates that did not exist, and no invitation had ever reached anyone.

Closing the loop, for real

One typed pipeline now carries 63 notification events to five channels (in-app, email, push, Slack and webhook), each behind its own preference check and idempotency key. Email leaves through a leased queue in batches of 100, paced under the provider's limit. On staging, one email now takes a median of 6 seconds end to end, down from about 100, and a burst of 100 was accepted in 3.1 seconds.

Letting Claude, ChatGPT or Cursor act inside a workspace is only safe if an agent can never do more than the person who connected it, whatever the model says.

AI agents that cannot overstep

Looph's MCP server (Model Context Protocol, the open standard AI assistants use to call tools) is a separate app with its own OAuth sign-in and consent screen, where the person chooses what the agent may do. For every request it mints a sixty-second database token for that person, so their own row-level security decides every row, and four more layers in the database can only take permissions away. It offers 28 tools (12 read, 16 write), logs and attributes every action, and the owner can switch agents off at any time. In production since 29 September 2026.

The operator area, which can see and change every customer's workspace, lived inside the same app customers use.

The operator console, out of the customer's app

Looph's operator console was rebuilt as a separate app on a host of its own, importing nothing from the product. Who you are is proved on your own token before the key that reaches every workspace is touched, operators hold one of three roles, and nobody can remove their own access or the last super-admin's.

At handover no change could be verified: type checking had never run, lint reported 19,849 problems so a real one was invisible, and two tests failed on untouched code.

Gates that tell the truth

Every gate now runs in CI against a recorded baseline, and a count that goes up fails the build: old debt is tolerated, new debt is refused. An architecture check fails the build when one layer reaches into another. Passing tests went from 1,152 at handover to 10,387 across five suites, with every test file green in CI on 29 September 2026.