Skip to main content
Glama
WYRE-AI
by WYRE-AI
README.md
# nutanix-mcp

Multitenant Streamable HTTP bridge over [nutanix/ntnx-api-mcp-server](https://github.com/nutanix/ntnx-api-mcp-server) — Nutanix's official Prism Central v4 API MCP server — built so the WYRE conduit gateway can forward per-tenant Nutanix credentials as HTTP headers.

> **Upstream pin:** Nutanix's server is a **technical preview**, pinned here at tag **`v0.8`** (Apache-2.0). See [Bumping the upstream pin](#bumping-the-upstream-pin).

## Why

The upstream server is **stdio-only** (`nutanix-mcp serve-stdio` is its only serve mode) and reads its Prism Central credentials from environment variables at process startup — single-tenant per process. Our gateway is multi-tenant: every request carries the calling org's credentials as HTTP headers, and the vendor container has to translate those headers into something the upstream understands.

Because the upstream has no HTTP mode to proxy to, this bridge holds an **MCP client session over stdio** to each tenant child and re-serves it over Streamable HTTP:

1. Listens on `:8080` with `POST /mcp` and `GET /health`. Startup requires a real `CONDUIT_S2S_SECRET`. Every `/mcp` request must carry a valid `X-Gateway-S2S` signature of that secret.
2. 401-gates every `/mcp` request on the `X-Nutanix-Pc-*` credential headers (below). Those headers are the only Prism Central credential source.
3. Lazily spawns one `nutanix-mcp serve-stdio` child per credential tuple (keyed by a hash), with the tenant's `PC_*` env vars set, and connects an MCP client to it over stdio.
4. Serves both protocol eras on `/mcp` via the v2 SDK's `createMcpHandler(factory, { legacy: 'stateless' })` — 2025-era `initialize`-handshake clients (the conduit gateway today) and modern 2026-07-28 envelope clients. `tools/list` and `tools/call` delegate to the tenant's child session.
5. Evicts idle children after 60 minutes (`IDLE_EVICT_MS`).

Tool names **pass through unchanged**: the upstream's 24 tools — 20 `{namespace}_execute` tools (aiops, clustermgmt, datapolicies, dataprotection, files, iam, licensing, lifecycle, microseg, monitoring, multidomain, networking, objects, opsmgmt, prism, security, storage, tenancy, vmm, volumes) plus 4 discovery tools (`listOperations`, `getOperationSchema`, `getCodeSample`, `getOperationPermissions`).

## Read-only in v1 — deliberate

Every child is spawned with **`READ_ONLY_MODE=true`** (also the upstream default): the upstream rejects all non-GET operations before they ever reach Prism Central. v1 of this bridge ships read-only as a deliberate fleet decision. Write support would be a reviewed, versioned change to `credentialsToChildEnv()` in `src/credentials.ts` — not a config flip.

## Credential contract

The gateway forwards these headers on every `/mcp` request; the bridge maps them onto the upstream child's environment. **conduit's vendor-config must match this table exactly.**

| Header | Child env var | Required | Notes |
|---|---|---|---|
| `X-Nutanix-Pc-Host` | `PC_HOST` | yes | Prism Central IP or FQDN |
| `X-Nutanix-Pc-Port` | `PC_PORT` | no | Upstream default `9440`. Upstream quirk: any port other than 9440 makes it use `http://`, not `https://` |
| `X-Nutanix-Pc-Username` | `PC_USERNAME` | with password | Basic auth pair |
| `X-Nutanix-Pc-Password` | `PC_PASSWORD` | with username | Basic auth pair |
| `X-Nutanix-Pc-Api-Key` | `PC_API_KEY` | alternative | Sent to PC as the `X-ntnx-api-key` request header; upstream prefers it over basic auth when both are set |
| `X-Nutanix-Pc-Insecure` | `PC_INSECURE` | no | `"true"`/`"false"` — skip TLS verification (default `false`) |

**Validity rule:** `pcHost` present AND (`apiKey` present OR `username`+`password` present). Anything else → **HTTP 401** with a JSON-RPC error body.

`READ_ONLY_MODE=true` is additionally forced on every child (see above).

## API-spec artifacts (baked at build time)

The upstream builds its tool surface from YAML API-spec artifacts, not from the live PC. `nutanix-mcp init` downloads them — and **without** PC credentials it runs in `latest_release` mode against the public `developers.nutanix.com` namespace API (no PC access required; verified empirically: 20 namespaces). The Docker build runs `init` once and bakes the artifacts into the image at `/opt/nutanix-mcp/artifacts`, shared read-only by all tenant children. Consequences:

- `tools/list` and the discovery tools work with **no reachable PC** — only `{namespace}_execute` calls touch Prism Central.
- Artifact versions are the latest public release at image build time, not the tenant PC's exact versions (the upstream's `pc_compatible` mode would need live PC access at spawn). For the read-only v1 surface this is the right trade: shared artifacts, fast tenant spawns.

## Configuration

Copy [`.env.example`](.env.example). Leave `CONDUIT_S2S_SECRET` empty in that file and set a real value only in the environment you run.

| Env var | Default | Notes |
|---|---|---|
| `CONDUIT_S2S_SECRET` | empty (refuses to start) | Service-to-service secret. The HTTP server logs an error and exits non-zero when this is empty or a known placeholder (`changeme`, `secret`, `replace-me`, `your-s2s-secret`, `replace-with-a-real-secret`, and the other example values). Generate one with `openssl rand -hex 32`. Compose will not start without it. The value is never logged. |
| `MCP_ALLOW_INSECURE_DEV` | unset | Set to exactly `1` only for local development to start HTTP without `CONDUIT_S2S_SECRET`. The process logs a warning, skips `X-Gateway-S2S`, and binds `127.0.0.1` only. A non-loopback `MCP_HTTP_HOST` refuses to start. A known placeholder secret still refuses to start. |
| `MCP_HTTP_HOST` | `0.0.0.0` when a real secret is set | Bind address. A configured secret honors this value, including when it is unset (historical `0.0.0.0` container bind). The image and Compose set `0.0.0.0` inside the container. With no secret, `MCP_ALLOW_INSECURE_DEV=1` binds `127.0.0.1` only. |
| `PORT` | `8080` | Public listen port (`MCP_HTTP_PORT` is the fallback). |
| `NUTANIX_MCP_DIR` | `/opt/nutanix-mcp` | Upstream checkout (venv + artifacts). |
| `NUTANIX_MCP_BIN` | `$NUTANIX_MCP_DIR/.venv/bin/nutanix-mcp` | Upstream console script the bridge spawns. |
| `ARTIFACTS_DIR` | `$NUTANIX_MCP_DIR/artifacts` | Baked YAML API-spec artifacts. |
| `CHILD_LOG_DIR` | `/tmp/nutanix-mcp-logs` | Writable dir for upstream per-process log files. |
| `IDLE_EVICT_MS` | `3600000` | Idle tenant timeout (60 min). |
| `SPAWN_TIMEOUT_MS` | `60000` | Max wait for a child to answer the MCP handshake. |

`/health` and `/healthz` stay unauthenticated and do not read credentials.

### HTTP authentication

With `CONDUIT_S2S_SECRET` set, every `/mcp` request must include `X-Gateway-S2S`. The check is `verifyS2sHeader` in [`src/s2s-verify.ts`](src/s2s-verify.ts): the header value is `t=<unix seconds>,v1=<hex>`, where the hex is HMAC-SHA256 of the literal string `t=<unix seconds>` keyed with `CONDUIT_S2S_SECRET` (64 lowercase hex characters). The timestamp must be within 300 seconds of the server clock. A missing or invalid header is HTTP 401, before Prism Central credentials are read.

Sign with the same `CONDUIT_S2S_SECRET` the server process is using. [Local development](#local-development) exports that value in the shell; reuse it for this request instead of generating another one.

```bash
TS=$(date +%s)
SIG=$(printf 't=%s' "$TS" | openssl dgst -sha256 -hmac "$CONDUIT_S2S_SECRET" -hex | awk '{print $NF}')
curl -sS http://127.0.0.1:8080/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -H "X-Gateway-S2S: t=${TS},v1=${SIG}" \
  -H 'X-Nutanix-Pc-Host: pc.example.com' \
  -H 'X-Nutanix-Pc-Api-Key: fake' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"dev","version":"0"}}}'
```

This bridge is gateway mode. `/mcp` also requires `X-Nutanix-Pc-Host` and either `X-Nutanix-Pc-Api-Key` or both `X-Nutanix-Pc-Username` and `X-Nutanix-Pc-Password`. Those headers are the only credential source for that request. `PC_*` values in the process environment are not used as a fallback.

## Local development

```bash
# 1. Get the upstream at the pinned tag with a venv + artifacts
git clone --branch v0.8 --depth 1 https://github.com/nutanix/ntnx-api-mcp-server ../ntnx-api-mcp-server
cd ../ntnx-api-mcp-server
uv venv .venv && uv pip install .
ARTIFACTS_DIR=$PWD/artifacts .venv/bin/nutanix-mcp init   # no PC creds needed
cd -

# 2. Build and run the bridge against it
npm ci && npm run build && npm test
# HTTP refuses to start until CONDUIT_S2S_SECRET is a real value.
# Export it so this shell keeps the same secret for the server, the
# signing request above, and the Docker commands below.
export CONDUIT_S2S_SECRET="$(openssl rand -hex 32)"
NUTANIX_MCP_DIR=../ntnx-api-mcp-server node dist/index.js

# Alternative: local HTTP only, without a service-to-service secret (always
# binds 127.0.0.1). A non-loopback MCP_HTTP_HOST makes this exit. The empty
# assignment applies only to this process; the exported secret stays in the shell.
MCP_ALLOW_INSECURE_DEV=1 CONDUIT_S2S_SECRET= \
  NUTANIX_MCP_DIR=../ntnx-api-mcp-server node dist/index.js

# 3. Smoke it (health is unauthenticated; /mcp needs X-Gateway-S2S — see above)
curl -s localhost:8080/health
```

Or build the image (it bakes everything, including a stdio smoke test that fails the build if `serve-stdio` can't answer `tools/list`). The image listens on `0.0.0.0` inside the container and still refuses to start without a real `CONDUIT_S2S_SECRET`. Publish the port on loopback. Prism Central credentials arrive on each request as `X-Nutanix-Pc-*` headers.

```bash
docker build --platform linux/amd64 -t ghcr.io/wyre-ai/nutanix-mcp:dev .
# Pass the exported secret. An empty value makes the container exit.
docker run --rm \
  -e CONDUIT_S2S_SECRET="$CONDUIT_S2S_SECRET" \
  -p 127.0.0.1:8080:8080 \
  ghcr.io/wyre-ai/nutanix-mcp:dev
```

Compose refuses to start when `CONDUIT_S2S_SECRET` is unset. The gateway service publishes `127.0.0.1:8080:8080` and does not inject `PC_HOST`, `PC_USERNAME`, `PC_PASSWORD`, `PC_API_KEY`, or `env_file`. Use the exported secret so a later signing request matches the container:

```bash
docker compose up
```

## Bumping the upstream pin

The upstream is pinned to the reviewed tag `v0.8` in the `Dockerfile` (`NUTANIX_MCP_REF`) — never `main` (NSA MCP guidance / fleet security baseline). To bump:

1. Review the upstream diff between the current pin and the new tag (tool surface, credential handling, `READ_ONLY_MODE` semantics).
2. Change `NUTANIX_MCP_REF` in the `Dockerfile` and the tag in this README.
3. `docker build` locally — the build-time smoke tests assert the venv entrypoint runs and the stdio tool surface still answers (update the expected tool count if namespaces changed).
4. Land as a `feat:`/`fix:` PR so semantic-release cuts a version.

## License

Apache-2.0. The bundled `ntnx-api-mcp-server` is Apache-2.0 by Nutanix.