# NotebookLM Expert Research MCP for Perplexity

Independent remote MCP for one authorized owner. It exposes a compact, read-oriented NotebookLM surface over authenticated Streamable HTTP and keeps its browser/auth profile, metadata, token, runtime and service separate from the Claude Desktop product.

## Product boundary

- `PRODUCT_MODE=PRIVATE_SINGLE_OWNER`
- One authorized Google identity and one owner-controlled remote MCP client.
- No multi-user/SaaS account model.
- Notebook/source bodies are returned only to the authenticated caller and are not persisted or logged.
- The NotebookLM adapter is an explicit `HIGH_EXTERNAL_UI_DEPENDENCY`; v1 prefers the upstream internal RPC path and fails closed when its contract or UI fallback drifts.

## Local verification

```powershell
npm.cmd install
npm.cmd run certify:local
```

The automated suite uses `TEST_ONLY_FAKE_ADAPTER` only under `NODE_ENV=test`. These checks certify server behavior, isolation and MCP protocol mechanics; they are not evidence of live NotebookLM authentication or research.

## Staging runtime

Copy `.env.example` to a root-readable staging environment file, generate a unique random token of at least 32 characters, and keep the service bound to `127.0.0.1`. Use the artifacts in `deployment/` for the dedicated Unix user, systemd service and nginx HTTPS reverse proxy.

The one-time legitimate Google ceremony uses the loopback-only noVNC scripts documented in `docs/AUTH_BOOTSTRAP.md`. Its auth command is:

```bash
sudo -u nlm-perplexity-stage /opt/nlm-expert-research-mcp-perplexity-staging/deployment/auth-bootstrap-staging.sh
```

It must run only through a secure interactive display method described in `docs/AUTH_BOOTSTRAP.md`. Never paste passwords, 2FA codes or cookies into chat.

## Tool policy

- `nlm_consultar`: ordinary targeted question.
- `nlm_investigar` `STANDARD`: bounded synthesis.
- `nlm_investigar` `DEEP`: bounded primary plus counter-evidence/gap pass when the measured remaining budget permits it.
- `nlm_verificar_respuesta`: NotebookLM source-grounded model evaluation, not deterministic proof.
- Source title scoping in v1 is disclosed as prompt-scoped, not a strict transport-level source filter.

See `docs/OPERATIONS.md`, `docs/AUTH_BOOTSTRAP.md`, `docs/UI_DRIFT_RUNBOOK.md`, and `docs/PERPLEXITY_ACCEPTANCE.md`.