Agents · Developer agents

Govern the coding agents on your developers’ machines.

Coding agents read files, run commands and call tools with a developer’s access. PRAE decides what they can execute. For Claude Code, a PreToolUse hook judges each built-in and MCP tool call before it runs. For Cursor, Windsurf and Claude Code, it reads the instruction files they follow and gates the file access they route through its MCP server.

Govern → Enforce → Prove

In part

01

Govern

One pack for every coding agent: the paths, hosts, commands and budgets a developer’s agent may use. The instructions it already follows — AGENTS.md, CLAUDE.md, .cursorrules, .windsurfrules — are read and proposed as rules, never applied unreviewed.
02

Enforce

For Claude Code, every tool call passes the hook first: refused, held as a permission prompt, or handed back to Claude Code. For Cursor and Windsurf, file reads routed through PRAE’s MCP server are inspected before their contents return.
03

Prove

Every decision sealed with the clause that produced it, so what a coding agent did on a developer’s machine is on the record.

Coverage today: Claude Code through a PreToolUse hook (prae-cli 0.10.0) — a shell command is judged by its text, not its effect, files attached with @ fire no hook, tool output is not screened, and what a person answers to a hold is not recorded. Cursor and Windsurf through PRAE’s MCP server for file access and through instruction ingest; their own built-in tools are not intercepted yet — that is on the roadmap. PRAE does not scan machines or discover agents.

Cursor →Windsurf →MCP & MCP servers →AI compliance & audit readiness →