What happened
On Oct 1, 2026 Tuskira released the Tuskira AI Agent Gateway on GitHub (github.com/Tuskira/tusk-ai-secured-gateway) under Apache-2.0, pre-1.0 alpha. The single self-hosted Go binary sits in front of LLM providers (Anthropic/OpenAI/Gemini/Bedrock) and MCP tool servers, authenticating every request, pulling real credentials from an encrypted store only at approved-call time (so agents never hold API keys/tokens), enforcing tool scoping per agent profile on every tools/call, and logging all activity; it also blocks egress to cloud-metadata/private addresses (SSRF) by default.
Why it matters
Credential sprawl around AI coding agents (Claude Code, Cursor, Codex) and MCP servers is a fast-emerging gap: the NSA and AWS have both flagged that MCP lacks identity/role enforcement and 'you must assume an agent can do anything within its entitlements.' A free, self-hosted, auditable gateway that centralizes credential injection and per-agent MCP policy gives platform/security teams a low-cost runtime control before commercial agent-gateway pricing.
Applicability
DevSecOps, platform engineering, and MCP-heavy enterprises (self-hosted/air-gapped too) should evaluate against commercial agent gateways for locked-down agent-to-tool access; relevant immediately for teams running Claude Code/Cursor/Codex in production.