Ai Agent Mcp Laboratories

---
name: ai-agent-mcp-laboratories
description: Run synthetic local agent and MCP boundary cases and interpret observed fixture effects.
---

# Local agent and MCP laboratories

Use this skill for offline rehearsal of prompt injection and tool authorization boundaries in the Foundry Agent Harness. The agent is a deterministic fixture, not a model or a Microsoft Foundry integration. The MCP server is a fixed local stdio fixture with synthetic state and canary data. Results describe these fixtures only.

## Run

```bash
python3 plugins/offensive-security/foundry-agent-harness/foundryharness/cli.py lab --json --output /tmp/cops-agent-mcp-lab-report.json
```

Omit `--json` for a concise summary. `--cases` accepts one case file or a directory of case files conforming to `schemas/lab-case.schema.json`. Use synthetic data only; case text and call arguments appear in reports. Do not add credentials, private data, or live endpoints.

Inspect `environment_verified`, `cleanup_verified`, and `expectation_met` for every result. Read `model_text`, `attempted_calls`, `denied_calls`, `executed_calls`, and `executed_effects` as separate evidence. An attempted call is not an effect. An executed call can be a read with no effect. Effects are determined from MCP state before and after execution. The baseline `permission_only` case deliberately permits a synthetic mutation; the paired `strict` case denies it.

The runner starts only its bundled Python fixture processes. It does not run user-selected commands, use hosted models or services, or expand authorization to live assessment. Its subprocess boundary is not an OS network sandbox. See `docs/PLAYBOOK.md` for interpretation and `specs/agent-mcp-laboratories.spec.md` for the contract.