Cinch is an MCP server + cross-harness skill set for building and operating AI agents safely. 117 enforceable controls, each mapped to a threat, a verification step, and a framework — queryable by any MCP-compatible agent at runtime.
The agent never trusts a prompt to enforce a boundary. It asks Cinch for the control, the verification step, and the framework mapping — and the harness enforces the answer.
Six checklists — agent containment, harness engineering, system hardening, red team, supply chain, agent environment. Each control names a threat, a required control, and how to verify it.
Every control maps to NIST AI RMF, OWASP LLM/Agentic, CUSTODY, LASM, and MITRE ATLAS — so a passing assessment is directly evidence for the obligations you already track.
Seven tools your agent queries live — checklist_run, threat_search, mapping_lookup, protocol_get, checklist_diff. The harness enforces; Cinch answers. The server is read-only by design.
Every answer becomes evidence. Export a PDF report, framework crosswalk appendix, signed JSON for CI gating, and a CSV gap register — re-assessment diffs are machine-readable.
CUSTODY addresses capability accretion: the gap between what an agent is granted and what it can actually do. Seven pillars define what an agent may access, change, invoke, or delegate.
Cinch runs as a local MCP server over stdio. Any MCP-compatible harness — Claude Code, Cursor, custom orchestrators — connects the same way.
pip install cinch-ai-security
cinch serveCopy
{
"mcpServers": {
"cinch": {
"command": "cinch",
"args": ["serve"]
}
}
}Copy
Cinch is the control plane sitting between the agent that wants to act and the harness that decides whether to let it.
Before a high-impact action, the agent queries Cinch: threat_search("prompt injection") or checklist_run("agent-containment").
Cinch returns the required control, the verification step, and the framework mapping — read-only, validated, path-traversal-hardened. No agent can write to it.
The harness checks the answer and gates the action. Evidence is logged. Re-assessment diffs against the last run for CI gating.
The interactive console below is a live assessment against the agent-containment checklist — mark controls enforced, log evidence, and watch the containment score and framework coverage update in real time.