Your credentials. Machine-speed decisions. Put request-level policy, human approvals, and audit between AI coding agents and your infrastructure.
AI coding agents such as Claude Code, Codex, and Cursor run infrastructure tools with your credentials. A mistaken instruction can become a production change at machine speed, before you have time to review it.
Prompt injection adds another route to the same permissions: untrusted repository content or tool output can steer an AI coding agent towards an action you never intended. Request-level policy checks the operation before it reaches the upstream API, independently of the AI coding agent's reasoning.
msh local supports kubectl, gcloud, aws, fly, cloudflare, and custom CLIs. Configure the tools to use the local proxy and select the hosts to inspect in policy. Interception is selective, rather than system-wide packet capture; traffic outside that configured path is outside its enforcement boundary.
The msh local daemon runs on your workstation and intercepts HTTPS requests for policy-listed hosts. Each inspected request is classified into one of four tiers, then evaluated against YAML policy. Policy can allow, deny, or hold a request for human approval before forwarding it.
Read-only operations, such as listing resources and checking status. Policy decides which reads may proceed.
Write operations that create or update resources. Require review where a change needs human oversight.
Destructive operations and access to secrets. Apply stricter policy to requests with a larger impact.
Administrative and high-risk operations, including permission changes and interactive access. Set explicit boundaries in policy.
Requests and decisions are recorded in an append-only JSONL audit log at ~/.mezite/local/audit.jsonl. The local file is not tamper-evident, so enrol workstations with MezHub to keep a central copy of every decision.
AI agents reach MCP servers through Mezite. Per-tool policy, approvals, and audit apply controls at the tool invocation boundary.
Review held local requests and approve or deny them in Mezite Desktop. The CLI alternative is msh local approval; requests awaiting approval stay blocked until a decision is made.
Enrol workstations with MezHub to receive centrally managed local policy. Administrators manage policy with mezctl local-policy get, set, and history.
When a workstation is enrolled, its local audit events are also uploaded to MezHub's audit log, giving the team a central record of requests and decisions.
Local request approvals are separate from time-boxed access escalation. Use Mezite Access Requests through msh requestwhen you need a temporary access grant.
Read AI coding agents changed the gamefor why request-level inspection now matters, then configure your local gateway or connect your MCP servers.