A REA call may read a file, add analysis state, launch a foreign program or change a connected device, depending on the operation. MCP annotations describe effects but are not permission controls. Keep agent access narrow, treat target content as untrusted and verify consequential conclusions from raw evidence before acting on a result.
Read the operation
The official MCP contract distinguishes target reads from operations that may run code or mutate a device. Inspect the connected tool list and host permissions before use. Begin with static inspection, then define an authorized target and isolated environment for any process or device action.
Defend the interpretation
Binary strings, web captures and pseudocode can contain instructions addressed to a model. They remain lower-trust sample data. Require addresses, hashes and tool output for consequential claims. Effect hints do not replace authorization or independent verification.
Requirements & evidence
- Resource type
- Guide
- Category
- Unknown
- Host platforms
- Unknown
- Targets
- native, android, web
- Level
- Intermediate
- Source checked
- 2026-11-04T10:07:15.376059+00:00
Primary sources
- https://raw.githubusercontent.com/morluto/rea/main/docs/mcp-contracts.md
- https://arxiv.org/abs/2605.30667
Source descriptions establish documented scope. They do not establish that every workflow has been reproduced by AI Craft Pad.