A brake on your agents that only you control: verifiable, no third party, no data leaves your machine for AgentGuard. Each post uses a fresh install and real output. A green verdict means the integration can stop the call before dispatch. Partial and no verdicts state exactly where the host leaves a bypass.
Reproduce how the local registry classifies one model ID across four routes, and see exactly what a signed provenance record does and does not prove.
guarded provider clientThe client wrapper stopped spend-cap and capability violations before network dispatch. The allowed call produced a signed receipt, and the four-entry chain verified offline.
guarded client before proxyA real local LiteLLM proxy saw one allowed call and neither refusal. Direct access to the proxy remains a bypass, so ingress must be restricted to guarded callers.
mandatory local hookThe pre-tool hook really blocks estimated spend and named destructive actions. Claude Code does not expose exact per-tool token cost here, and the integration does not enforce a generic capability tier.
MCP decision toolThe MCP server returns real signed spend decisions, but Cursor is not forced to consult it and the current tool has no capability input. This post shows the failed probe.
MCP extensionThe MCP tool can return signed policy decisions to Goose, but MCP alone is not an unavoidable call-path interposer. Treat the older setup as agent-consulted decision support.
buzz-acp harnessThe harness exposes AgentGuard decisions through MCP. It is useful signed policy evidence, but the current MCP server is not by itself a mandatory interposer.
Every post answers the same question the same way, so a new one is a short job, not a project:
1. Identify the exact interception point. 2. Install the published package in a fresh scratch directory. 3. Capture an allowed signed decision, a spend refusal, and a capability refusal. 4. Verify offline and reject a tamper. If the host cannot produce one of those outcomes, the badge and post say so.