PRAE

Integrations · Agent surfaces · Agent surface

PRAE for Claude Desktop, through MCP.

In part

Claude Desktop reaches local files and tools through the MCP servers it mounts. PRAE’s MCP server speaks standard MCP over stdio, so Claude Desktop can use it as its file tool, with every read judged first.

Serves: Developer agents →Custom / homegrown agents →

  1. 01 · Where the agent operates

    Claude Desktop

  2. 02 · What it can reach

    Whatever its MCP servers expose: local files, and the tools and services they connect to.

  3. 03 · Where PRAE intervenes

    At PRAE’s MCP server: file reads routed through it are judged before contents return.

04 · What PRAE does: Evaluate → Enforce → Steer → Prove

  1. Evaluate

    Each file read the agent routes through PRAE’s MCP server is judged against your pack before it runs.

  2. Enforce

    A refused read returns nothing: the file’s contents never reach the model, and the agent is told why.

  3. Steer

    Not on this surface yet.

  4. Prove

    Every decision is sealed in the hash-chained, signed ledger, with the rule that produced it.

05 · What you get, and how it plugs in

  1. Register PRAE’s MCP server npx -p prae-cli prae mcp --root <workspace> --pack <policy.json> --ledger <decisions.jsonl>

Covers

File reads and listings routed through PRAE’s MCP server.

Limits

Only what goes through PRAE’s MCP server is seen; other MCP servers it mounts are not. Not yet validated end to end in Claude Desktop.