I built three open-source tools for coding agents: onsystem, ondocs and signoff-ui

Description and use

Построил три opensource-инструмента для когниных агентов: onsystem (гарантирует соответствие дизайна системы при написании React-кода агентами), ondocs (opensource-ассиринг документации с циментированными ответами и MCP-сервером для агентов) и signoff-ui

Problem

Coding agents now write a real share of the code in many teams, but three parts still go wrong often: (1) agents write UI from their training data instead of from the team's design system; (2) docs are written for people but visitors and coding agents need answers and an entry point; (3) someone must review what the agent changed and decide what it may do.

Outcome

In a 1,200-run benchmark across 6 models from 3 providers, onsystem 0.5.0 reduced design-system errors per run from 0.85 to 0.00 with no measurable change in tsc pass rate or cost. Scored by @shadcn/lint (not part of onsystem), findings per run went from 4.97 to 2.64.

Workflow

• Install onsystem once in the design-system package; every app in the monorepo inherits it.
• Run npx -y onsystem check . to check agent-written React files for design-system errors.
• onsystem flags palette colors, native elements and APIs the design system lacks, with a fix applied automatically.
• Install ondocs plugins (Starlight, Docusaurus, VitePress), guides (Next.js, Fumadocs, Nextra), or one script tag to enable cited AI answers from the team's own model key.
• Install signoff-ui React components for multi-file diff review and tool approvals inside assistant-ui, AI Elements, or a custom chat.
• Review agent changes via signoff-ui diffs and approve or reject specific tool calls.

Evidence

Keeps coding agents on your design system. It catches palette colors, native elements and APIs your design system lacks in agent-written React, at every edit and pull request, with the fix.

Source

I tested it in a 1,200-run benchmark across 6 models from 3 providers. Counted by onsystem 0.5.0's own check, design-system errors per run went from 0.85 without it to 0.00 with its skill, MCP server and hook, with no measurable change in tsc pass rate or cost.

Source

Scored by @shadcn/lint, a linter that is not part of onsystem, findings per run went from 4.97 to 2.64.

Source

Open-source Ask AI for your docs. Visitors get cited answers on your own model key, and 'I don't know' instead of a guess when the docs don't cover it. Coding agents get an MCP server from the same index.

Source

Review what your agent changed. Approve what it may do. React components for multi-file diff review, where the agent gets back exactly what you accepted, and tool approvals once, for the session or always.

Source