litellm-mcp-skills-gateway

LiteLLM MCP Skills Gateway

Workflow

1. Distinguish among four types of tasks: an agent using an MCP tool, directly calling the MCP REST API, publishing a skill, and managing LiteLLM Proxy.
2. Check the LiteLLM version, transport, auth, permissions, toolsets, and Skills Gateway schema.
3. Before registering an MCP server, review the source, authentication, data flow, available tools, and health status.
4. Establish a minimal toolset based on the intersection of key, team, organization, and agent permissions.
5. Pair high-risk tools with litellm-guardrails-safety to restrict input and execution behavior.
6. Pin external skills to a reviewed commit SHA, and record the content hash and required permissions.
7. Test list, allowed call, denied call, source update, and revocation in an isolated Proxy.

When supply chain rules, the Skills Gateway process, and experiment specifications are needed, read Experiments and References。

Security Baseline

• Do not expose all tools of an entire MCP server by default.
• Do not let students hold production Proxy admin keys.
• Skills that manage users, keys, models, or agents use least privilege and manual approval.
• Git branches or mutable tags are not considered reviewable production skill sources.

Acceptance

• The tools available to each agent can be traced to the key, team, organization, and agent configuration.
• Unauthorized tools are invisible and cannot be called.
• Skill source, commit SHA, hash, permissions, and upgrade records are complete.
• Tests do not perform writes to the production Proxy or production MCP data.