Brightwing Systems

Software with a hard problem in the middle of it.

A roguelike simulation that had to become server-authoritative before it could be multiplayer. Generated mysteries checked for solvability before they ship. A spend limit an SDK can verify locally, without asking us first. 23 builds — and the reason to open any of them is what was hard about it.

Start a conversation

Things we needed while shipping the rest of this list.

AICostManager

Cost governance that can stop a runaway job, not just draw it a chart.

Django · Postgres · Celery · MCP

AI Token Pricer

The pricing table we needed internally, published because everyone needs it.

Django · shares the AICostManager deployment

AI Prompt Bucket

Prompts treated as deployable artefacts, with environments and an audit trail.

Django · Postgres · Celery · MCP

ConfigPig

One config registry that hands each consumer the format it wants.

Django · Postgres · Celery · MCP

PatchworkMCP

Agents hit dead ends silently; this turns each one into a pull request.

Django · Postgres · Celery · drop-ins for 5 languages

MCP Scoreboard

Every public MCP server, discovered and graded without being asked.

Django · Postgres · sandboxed probes

MCP Manager

Open source

Every AI tool keeps its MCP config somewhere different. This is one control plane.

Rust · Tauri · open source

Systems, not forms

A Python simulation the browser only renders, so a shared run is possible at all. A generator with a validator in front of it, because an unsolvable mystery is worse than none. Procedural galaxies built from a seed rather than placed by hand. These were picked because they had a part that could fail.

Tests that encode intent

Every build starts from one Django baseline whose suite asserts why, not just what — contrast ratios computed against published floors in both themes, a guard against stale compiled CSS, a check that fails when a placeholder would reach production. Incidents get written down once instead of re-learned.

AI as a component

Characters that improvise rather than branch. Agents that report the gap they just worked around. Cost governance for teams shipping the same. Every model call goes through one facade with a named purpose, so a price change is one edit and not an audit.