MCP · OpenAI Codex
Deslint for Codex
OpenAI Codex is great at one-shotting whole features. Without a verification layer, every feature is also a fresh opportunity for arbitrary Tailwind values, off-palette colours, and missing ARIA attributes. Deslint's MCP server gives Codex a deterministic callable tool surface — analyze_and_fix, compliance_check, enforce_budget — it can invoke before committing.
Install
Any MCP-compatible Codex client accepts the standard MCP config. Add the deslint entry to your Codex MCP config file:
{
"mcpServers": {
"deslint": {
"command": "npx",
"args": ["-y", "@deslint/mcp"]
}
}
}If your Codex client hasn't enabled MCP yet, the deslint CLI runs the same rule set — run npx @deslint/cli scan in your Codex-generated diff to get the identical output.
What Codex gets from deslint
- Rule-backed structured output. Each tool response is a JSON payload Codex can parse without chain-of-thought guessing — rule id, line number, suggested fix, confidence.
- Determinism the planner can trust. Codex can assert a budget (“no new
no-arbitrary-colorsviolations”) and verify it mechanically withenforce_budget. No need to re-ask a separate reviewer model. - Compliance report on demand.
compliance_checkreturns the full WCAG + design-token scorecard in one call — useful when Codex is writing release notes or PR descriptions.
Example loop
Data boundary
Deslint's MCP server runs entirely on the host machine and does not call OpenAI, Anthropic, or any other provider. Even if your Codex client is cloud-hosted, deslint itself processes your code locally. For regulated teams running Codex on isolated infrastructure, deslint ships as a single stdio binary with no network dependency.