Summary
LiteLLM 1.102.0-rc.1 exposes registered Agent Skills through a well-known index and can configure Codex and Claude Code to use a gateway key. Its MCP gateway now applies per-user tool permissions to both discovery and execution, while OAuth handling and guardrail failure paths get additional hardening. The release also adds customer-managed KMS keys for virtual-key secrets and routing observability headers. It is installable from PyPI, but remains a release candidate rather than stable GA.
Why it matters
For teams standardizing multiple coding agents behind one gateway, this release joins skill discovery, MCP authorization, secret custody and routing telemetry in the same control plane. The shape is useful for a proof of concept, but production teams should wait for the stable build and test permission-list consistency, spend accounting and guardrail failure behavior during upgrades.
Technical details
| Release | v1.102.0-rc.1; first PyPI artifact uploaded 2026-09-13T03:58:55Z; GitHub release published 2026-09-13T04:56:38Z; prerelease |
|---|---|
| Agent Skills | serves registered skills through an Agent Skills well-known index and configures Claude Code or Codex with a gateway key |
| Mcp Controls | per-user MCP tool permissions apply to tools/list and tools/call; MCP OAuth and discovery handling improved; Conduct Guard integration added |
| Key Management | customer-managed KMS keys supported for virtual keys stored in AWS Secrets Manager |
| Observability | response headers expose the routed model, savings and complexity-routing decisions |
| Reliability | fixes cover spend double counting, Redis circuit-breaker state, streamed Bedrock spend buffering, secret redaction and fail-closed guardrail paths |
| Availability | installable release candidate on PyPI; stable latest remains 1.100.1 at scan cutoff |
Tags
litellmllm-gatewayagent-skillsmcpauthorizationobservability