LiteLLM Proxy documentation describes an MCP Gateway that exposes list-tools, call-tools, prompts, and resources operations through a fixed endpoint while enforcing access by API key, team, or organization. Supported transports listed on docs.litellm.ai include Streamable HTTP, SSE, and stdio; operators can register HTTP, SSE, or stdio MCP servers through the LiteLLM UI or config.yaml after enabling database storage (`store_model_in_db` / `STORE_MODEL_IN_DB`). Release notes cited in the docs state LiteLLM v1.80.18 aligns with MCP protocol version 2025-11-25 and namespaces tools by MCP server name per SEP-986 naming rules for newly added servers. The gateway is positioned as a way to use MCP tools alongside all LiteLLM-supported chat models from Cursor or other OpenAI-compatible clients pointed at the proxy.
Use cases
- Centralize dozens of MCP servers behind one LiteLLM virtual key for a platform team
- Apply team-level budgets and guardrails before agents invoke third-party MCP tools
- Register Streamable HTTP or SSE MCP endpoints through the admin UI instead of per-developer JSON files
- Namespace tools per MCP server to avoid collisions when multiple catalogs are enabled
- Troubleshoot connectivity using the documented MCP troubleshooting guide after enabling DB storage
Key features
- Cursor
- VS Code
- Claude Desktop
- Codex
Frequently Asked Questions
- Do MCP servers persist without database storage?
- Docs require `store_model_in_db` / `STORE_MODEL_IN_DB=True` (or equivalent config.yaml) to store MCP definitions in the proxy database.
- Which MCP protocol version does LiteLLM cite?
- Documentation references MCP protocol version 2025-11-25 starting in LiteLLM v1.80.18, with SEP-986-compliant server names for new registrations.
- Can I mix stdio and remote HTTP MCPs?
- Yes—the overview lists Streamable HTTP, SSE, and stdio transports with separate UI walkthroughs for each pattern.
Related
Related
3 Indexed items
E2B MCP Gateway
E2B documents an MCP gateway that runs inside cloud sandboxes, exposing 200+ tools from the Docker MCP Catalog (Browserbase, Exa, Notion, Stripe, GitHub, and others) through a unified HTTP endpoint with bearer-token auth. Developers create a Sandbox with an `mcp` configuration map of server credentials, call `getMcpUrl()` / `getMcpToken()`, and attach the gateway to MCP clients such as Claude Code via `claude mcp add --transport http`. Sandboxes provide an internet-connected Linux environment where agents can install packages, run terminal commands, and execute generated code while MCP tools stay type-safe per E2B's overview at e2b.dev/docs/mcp.
Composio MCP Server
Composio documents MCP server creation through its SDK and dashboard at docs.composio.dev: developers call `composio.mcp.create()` with toolkit names, auth config IDs, and an `allowed_tools` list, then generate per-user MCP URLs via `composio.mcp.generate(user_id, mcp_config_id)`. Hosted endpoints follow the pattern `https://backend.composio.dev/v3/mcp/{SERVER_ID}?user_id=...` and require an `x-api-key` header when `require_mcp_api_key` is enabled (default for new orgs). Docs show wiring these URLs into OpenAI Responses API, Anthropic MCP client beta, Mastra MCPClient, Claude Desktop, and Cursor. Composio notes that dynamic sessions are recommended for most use cases, while single-toolkit MCP configs suit fixed integration surfaces.
Mem0 MCP Server
Mem0 documents a hosted Model Context Protocol server at https://mcp.mem0.ai/mcp that exposes Platform memory tools (`add_memory`, `search_memories`, `get_memories`, `update_memory`, `delete_memory`, `delete_all_memories`, `delete_entities`, `list_entities`, `list_events`, `get_event_status`) to Claude, Claude Code, Codex, Cursor, Windsurf, VS Code, and OpenCode. Setup uses `npx mcp-add` with HTTP transport or manual JSON/TOML client configs; Codex requires `MEM0_API_KEY` as bearer token per docs.mem0.ai/platform/mem0-mcp. The cloud server needs a Mem0 Platform API key from the dashboard and Node.js for the installer—no local vector database required for the hosted path.