diff --git a/docs/my-website/docs/mcp_zero_trust.md b/docs/my-website/docs/mcp_zero_trust.md new file mode 100644 index 00000000000..8f431523cb8 --- /dev/null +++ b/docs/my-website/docs/mcp_zero_trust.md @@ -0,0 +1,294 @@ +import Tabs from '@theme/Tabs'; +import TabItem from '@theme/TabItem'; + +# MCP Zero Trust Auth (JWT Signer) + +![Zero Trust MCP Gateway](/img/mcp_zero_trust_gateway.png) + +MCP servers have no built-in way to verify that a request actually came through LiteLLM. Without this guardrail, any client that can reach your MCP server directly can call tools — bypassing your access controls entirely. + +`MCPJWTSigner` fixes this. It signs every outbound tool call with a short-lived RS256 JWT. Your MCP server verifies the signature against LiteLLM's public key. Requests that didn't go through LiteLLM have no valid signature and are rejected. + +--- + +## Basic setup + +Add the guardrail to your config and point your MCP server at LiteLLM's JWKS endpoint. Every tool call gets a signed JWT automatically — no changes needed on the client side. + +```yaml title="config.yaml" +mcp_servers: + - server_name: weather + url: http://localhost:8000/mcp + transport: http + +guardrails: + - guardrail_name: mcp-jwt-signer + litellm_params: + guardrail: mcp_jwt_signer + mode: pre_mcp_call + default_on: true + issuer: "https://my-litellm.example.com" # defaults to request base URL + audience: "mcp" # default: "mcp" + ttl_seconds: 300 # default: 300 +``` + +**Bring your own signing key** — recommended for production. Auto-generated keys are lost on restart. + +```bash +export MCP_JWT_SIGNING_KEY="-----BEGIN RSA PRIVATE KEY-----\n..." +# or point to a file +export MCP_JWT_SIGNING_KEY="file:///secrets/mcp-signing-key.pem" +``` + +**Build a verified MCP server with [FastMCP](https://gofastmcp.com):** + +```python title="weather_server.py" +from fastmcp import FastMCP, Context +from fastmcp.server.auth.providers.jwt import JWTVerifier + +auth = JWTVerifier( + jwks_uri="https://my-litellm.example.com/.well-known/jwks.json", + issuer="https://my-litellm.example.com", + audience="mcp", + algorithm="RS256", +) + +mcp = FastMCP("weather-server", auth=auth) + +@mcp.tool() +async def get_weather(city: str, ctx: Context) -> str: + caller = ctx.client_id # JWT `sub` — the verified user identity + return f"Weather in {city}: sunny, 72°F (requested by {caller})" + +if __name__ == "__main__": + mcp.run(transport="http", host="0.0.0.0", port=8000) +``` + +FastMCP fetches the JWKS automatically and re-fetches when the signing key changes. + +LiteLLM publishes OIDC discovery so MCP servers find the key without any manual configuration: + +``` +GET /.well-known/openid-configuration → { "jwks_uri": "https:///.well-known/jwks.json" } +GET /.well-known/jwks.json → { "keys": [{ "kty": "RSA", "alg": "RS256", ... }] } +``` + +> **Read further only if you need to:** thread a corporate IdP identity into the JWT, enforce specific claims on callers, add custom metadata, use AWS Bedrock AgentCore Gateway, or debug JWT rejections. + +--- + +## Thread IdP identity into MCP JWTs + +By default the outbound JWT `sub` is LiteLLM's internal `user_id`. If your users authenticate with Okta, Azure AD, or another IdP, the MCP server sees a LiteLLM-internal ID — not the user's email or employee ID. + +With verify+re-sign, LiteLLM validates the incoming IdP token first, then builds the outbound JWT using the real identity claims from that token. The MCP server gets the user's actual identity without ever having to trust the original IdP directly. + +```yaml title="config.yaml" +guardrails: + - guardrail_name: mcp-jwt-signer + litellm_params: + guardrail: mcp_jwt_signer + mode: pre_mcp_call + default_on: true + issuer: "https://my-litellm.example.com" + + # Validate the incoming Bearer token against the IdP + access_token_discovery_uri: "https://login.microsoftonline.com/{tenant}/v2.0/.well-known/openid-configuration" + verify_issuer: "https://login.microsoftonline.com/{tenant}/v2.0" + verify_audience: "api://my-app" + + # Which claim to use for `sub` in the outbound JWT — first non-empty value wins + end_user_claim_sources: + - "token:sub" # from the verified incoming JWT + - "token:email" # fallback to email + - "litellm:user_id" # last resort: LiteLLM's internal user_id +``` + +If the incoming token is **opaque** (not a JWT — some IdPs issue these), add an introspection endpoint. LiteLLM will POST the token to it (RFC 7662) and use the returned claims: + +```yaml + token_introspection_endpoint: "https://idp.example.com/oauth2/introspect" +``` + +**Supported `end_user_claim_sources` values:** + +| Source | Resolves to | +|--------|-------------| +| `token:` | Any claim from the verified incoming JWT (e.g. `token:sub`, `token:email`, `token:oid`) | +| `litellm:user_id` | LiteLLM's internal user ID | +| `litellm:email` | User email from LiteLLM auth context | +| `litellm:end_user_id` | End-user ID if set separately | +| `litellm:team_id` | Team ID from LiteLLM auth context | + +--- + +## Block callers missing required attributes + +Some MCP servers expose sensitive operations that should only be reachable by verified employees — not service accounts, not external API keys. You can enforce this at the LiteLLM layer so the MCP server never receives the request at all. + +`required_claims` rejects with `403` if the incoming token is missing any listed claim. `optional_claims` forwards claims that are useful but not mandatory. + +```yaml title="config.yaml" +guardrails: + - guardrail_name: mcp-jwt-signer + litellm_params: + guardrail: mcp_jwt_signer + mode: pre_mcp_call + default_on: true + + access_token_discovery_uri: "https://idp.example.com/.well-known/openid-configuration" + + # Service accounts without `employee_id` are blocked before the tool runs + required_claims: + - "sub" + - "employee_id" + + # Forward these into the outbound JWT when present — skipped silently if absent + optional_claims: + - "groups" + - "department" +``` + +**What the client sees when blocked:** +```json +HTTP 403 +{ "error": "MCPJWTSigner: incoming token is missing required claims: ['employee_id']. Configure the IdP to include these claims." } +``` + +--- + +## Add custom metadata to every JWT + +Your MCP server may need context that LiteLLM doesn't carry natively — which deployment sent the request, a tenant ID, an environment tag. Use claim operations to inject, override, or strip claims from the outbound JWT. + +```yaml title="config.yaml" +guardrails: + - guardrail_name: mcp-jwt-signer + litellm_params: + guardrail: mcp_jwt_signer + mode: pre_mcp_call + default_on: true + + # add: insert only when the key is not already in the JWT + add_claims: + deployment_id: "prod-us-east-1" + tenant_id: "acme-corp" + + # set: always override — even if the claim came from the incoming token + set_claims: + env: "production" + + # remove: strip claims the MCP server shouldn't see + remove_claims: + - "nbf" # some validators reject nbf; remove it if yours does +``` + +Operations run in order — `add_claims` → `set_claims` → `remove_claims`. `set_claims` always wins over `add_claims`; `remove_claims` beats both. + +--- + +## AWS Bedrock AgentCore Gateway + +Bedrock AgentCore Gateway uses two separate JWTs: one to authenticate the transport connection and another to authorize tool calls. They need different `aud` values and TTLs — a single JWT won't work for both. + +LiteLLM can issue both in one hook and inject them into separate headers: + +```yaml title="config.yaml" +guardrails: + - guardrail_name: mcp-jwt-signer + litellm_params: + guardrail: mcp_jwt_signer + mode: pre_mcp_call + default_on: true + issuer: "https://my-litellm.example.com" + audience: "mcp-resource" # for the MCP resource layer + ttl_seconds: 300 + + # Second JWT for the transport channel — same sub/act/scope, different aud + TTL + channel_token_audience: "bedrock-agentcore-gateway" + channel_token_ttl: 60 # transport tokens should be short-lived +``` + +LiteLLM injects two headers on every tool call: +- `Authorization: Bearer ` — audience `mcp-resource`, TTL 300s +- `x-mcp-channel-token: Bearer ` — audience `bedrock-agentcore-gateway`, TTL 60s + +Both tokens are signed with the same LiteLLM key, so your MCP server only needs to trust one JWKS endpoint. + +--- + +## Control which scopes go into the JWT + +By default LiteLLM generates least-privilege scopes per request: +- Tool call → `mcp:tools/call mcp:tools/{name}:call` +- List tools → `mcp:tools/call mcp:tools/list` + +If your MCP server does its own scope enforcement and needs a specific format, set `allowed_scopes` to replace auto-generation entirely: + +```yaml title="config.yaml" +guardrails: + - guardrail_name: mcp-jwt-signer + litellm_params: + guardrail: mcp_jwt_signer + mode: pre_mcp_call + default_on: true + + allowed_scopes: + - "mcp:tools/call" + - "mcp:tools/list" + - "mcp:admin" +``` + +Every JWT carries exactly those scopes regardless of which tool is being called. + +--- + +## Debug JWT rejections + +Your MCP server is returning 401 and you're not sure what's in the JWT. Enable `debug_headers` and LiteLLM adds a `x-litellm-mcp-debug` response header with the key claims that were signed: + +```yaml title="config.yaml" +guardrails: + - guardrail_name: mcp-jwt-signer + litellm_params: + guardrail: mcp_jwt_signer + mode: pre_mcp_call + default_on: true + debug_headers: true +``` + +Response header: +``` +x-litellm-mcp-debug: v=1; kid=a3f1b2c4d5e6f708; sub=alice@corp.com; iss=https://my-litellm.example.com; exp=1712345678; scope=mcp:tools/call mcp:tools/get_weather:call +``` + +Check that `kid` matches what the MCP server fetched from JWKS, `iss`/`aud` match your server's expected values, and `exp` hasn't passed. Disable in production — the header leaks claim metadata. + +--- + +## JWT claims reference + +| Claim | Value | +|-------|-------| +| `iss` | `issuer` config value (or request base URL) | +| `aud` | `audience` config value (default: `"mcp"`) | +| `sub` | Resolved via `end_user_claim_sources` (default: `user_id` → api-key hash → `"litellm-proxy"`) | +| `act.sub` | `team_id` → `org_id` → `"litellm-proxy"` (RFC 8693 delegation) | +| `email` | `user_email` from LiteLLM auth context (when available) | +| `scope` | Auto-generated per tool call, or `allowed_scopes` when set | +| `iat`, `exp`, `nbf` | Standard timing claims (RFC 7519) | + +--- + +## Limitations + +- **OpenAPI-backed MCP servers** (`spec_path` set) do not support JWT injection. LiteLLM logs a warning and skips the header. Use SSE/HTTP transport servers to get full JWT injection. +- The keypair is **in-memory by default** and rotated on each restart unless `MCP_JWT_SIGNING_KEY` is set. FastMCP's `JWTVerifier` handles key rotation transparently via JWKS key ID matching. + +--- + +## Related + +- [MCP Guardrails](./mcp_guardrail) — PII masking and blocking for MCP calls +- [MCP OAuth](./mcp_oauth) — upstream OAuth2 for MCP server access +- [MCP AWS SigV4](./mcp_aws_sigv4) — AWS-signed requests to MCP servers diff --git a/docs/my-website/img/mcp_zero_trust_gateway.png b/docs/my-website/img/mcp_zero_trust_gateway.png new file mode 100644 index 00000000000..3955cef0553 Binary files /dev/null and b/docs/my-website/img/mcp_zero_trust_gateway.png differ