Coder AI Governance (AI Gateway and Agent Firewall)

door Coder

AI Tool

Laag vertrouwen

Coder AI Governance is a self-hosted control layer for AI coding tools. The AI Gateway authenticates model requests to a named user and logs prompts, token usage, models and tool invocations centrally, keeps LLM provider API keys in the control plane, and applies spend limits; the Agent Firewall wraps agent processes in a default-deny HTTP and HTTPS allowlist (domains, methods and paths), blocks other outbound protocols, and logs every allow and deny; audit events can be exported to a SIEM, and identity comes from OIDC and SCIM. It works with Coder Agents and with bring-your-own agents such as Claude Code and Codex. Rated medium: it is a risk-reducing control rather than an agent, but it becomes a concentrated store of every prompt, code snippet and tool call plus the organization's LLM provider keys, so its own security and retention settings matter a great deal, and it sits on the same platform whose registry was compromised on 31 Aug 2026 (advisory GHSA-vx42-ghc9-gw65). The firewall restricts HTTP and HTTPS egress and does not by itself address what an agent does inside an allowed domain. Coder says it has no access to customer workspaces or source code, and the data stays in the customer's deployment; it also collects product telemetry by default, and its Privacy Policy (free users cannot opt out) and the AI Governance FAQ (can be turned off) differ on this. No documentation of what the gateway logs, for how long, or how access to those logs is controlled was read, and the Trust Center's AI section is a one-line statement.