@cross-deck/ai

Prism — Crossdeck's intelligence layer for AI

Ask your app anything — revenue, errors, database read-cost, growth — in plain English, and get rendered charts and dashboards back.

"Which paying customer did this crash hit, and how much revenue is at risk?" — one question, answered across error, identity, and revenue. No single-layer tool can do that.

Why Prism is different

Most MCP servers are a connector to one system. Prism is an intelligence layer over your whole app — because Crossdeck already joins the layers about your users (identity · revenue · entitlements · errors · analytics · database read-cost) into one source of truth, by identity.

The point isn't "let AI read my data." It's the crossing: one question can span layers no other tool joins —

"This error a1b2c3 — who did it hit, and how many of them pay us?" → the named paying users behind the crash, each with their monthly revenue at risk, sorted by what's on the line.

That's the moat sentence. Prism gives it a voice in Claude, ChatGPT, Cursor, or any MCP client.

Permissions & data access

Secure, read-only access to your live Crossdeck workspace. Prism reads; it never writes.

Prism can see (the layers Crossdeck already owns about your app):

💷 Revenue	MRR, paying-customer counts, per-rail split
🐛 Errors	issues, blast radius, who was affected
📈 Analytics	per-host page views, unique visitors, top pages/referrers
🗄️ Read-cost	per-user-vs-overhead database reads, by operation
🧭 Customer journeys	the cross-layer view of one customer (pay × entitlements × cost)

Row-level customer identity is returned only when your connection was explicitly granted that scope — otherwise Prism answers from aggregates and tells you so. That boundary is enforced by Crossdeck, not left to the model.

Prism cannot create, modify, delete, or configure anything in your projects. There is no write path — every tool is read-only (readOnlyHint: true). Ask it to change something and it will tell you it can't.

Data is fetched live at question time over an OAuth-protected endpoint; the connector stores nothing.

Connect in 60 seconds

Remote (recommended) — one-click OAuth, no keys to paste

Prism is a hosted, OAuth-protected MCP endpoint:

https://api.cross-deck.com/mcp

Add it as a connector in your client, sign in with your Crossdeck account, pick a project, and approve. That's it — the client receives a short-lived, scoped workspace credential automatically.

• Claude (web / desktop): Settings → Connectors → Add custom connector → paste the URL above.
• ChatGPT: Settings → Connectors → Add → paste the URL.
• Cursor: Settings → MCP → Add → paste the URL.

Or open your dashboard at app.cross-deck.com → Developers → Prism → Connect AI for one-click "Add to Claude / ChatGPT / Cursor".

Local (Claude Desktop / Cursor / Claude Code) — stdio + a workspace key

No install step — npx fetches the package and its dependencies on each launch, so it self-heals and stays current (nothing local to go stale):

{
  "mcpServers": {
    "crossdeck": {
      "command": "npx",
      "args": ["-y", "@cross-deck/ai"],
      "env": { "CROSSDECK_SECRET_KEY": "cd_wk_live_…" }
    }
  }
}

Use a workspace key (cd_wk_) from your dashboard → API keys — one key reads your whole portfolio (every app), with identity, which is what Prism is for. A single-app project key (cd_sk_) also works in the same CROSSDECK_SECRET_KEY slot but sees only that one app. Never a publishable (cd_pub_) one. CROSSDECK_API_BASE overrides the endpoint (e.g. sandbox).

If the connector shows "failed" / "Server disconnected" — two known causes, both fixable:

• spawn npx ENOENT / command not found. Desktop clients launch with a minimal PATH that often excludes your Node bin (common with nvm). Fix: use the absolute path to npx as command — run which npx and paste the full path (e.g. /Users/you/.nvm/versions/node/vX/bin/npx).
• ERR_MODULE_NOT_FOUND … @modelcontextprotocol/sdk. This only happens if args points at a hand-built local checkout (…/dist/server.js) — that folder's node_modules can be cleaned out from under you (e.g. anything under /tmp). Fix: never point at a local build. Let npx -y @cross-deck/ai (above) resolve the published package, so its dependencies are always present.

Simplest of all: use the remote connector at the top of this section — it runs no local Node at all, so neither failure can occur.

Try these first

Ask	Tool	You get
"What's our MRR and paying-customer count, split by Stripe/Apple/Google?"	get_revenue	MRR, paying count, three-rail split, optional daily trend
"This error a1b2c3 — who did it hit and do any of them pay us?"	get_error_affected_users	The named paying users behind the crash, each with monthly revenue at risk
"Show me everything about customer agent_8842."	open_moat_dashboard	A rendered dashboard: what they pay × entitlements × read-cost, joined by identity
"Draw our user growth over the last 30 days."	draw_user_growth	A rendered interactive line chart of visitors + page views
"What's driving our database reads — per-user vs overhead?"	get_read_cost	The per-user-vs-overhead split + reads by operation

Data tools return clean JSON; draw_user_growth and open_moat_dashboard render interactive charts/dashboards inline in hosts that support MCP Apps (and fall back to a text summary elsewhere).

Tools

Thirteen tools — all read-only (readOnlyHint), each with a human-readable title, scoped/paginated output, and actionable errors. Every read is a point-read of a maintained ledger, so asking questions never runs up your database bill.

Every number is self-describing: values come back with a meta.semantics briefing and a coverage state, so a metric is reported as exactly what it is, and a blind spot (not_instrumented) is never mistaken for a real zero.

Portfolio

• list_projects — list every app this connector can read (your portfolio).
• get_portfolio — the coverage map: for each app, which surfaces are instrumented vs a blind spot. Load it first so a later blank reads as a known blind spot, not a fact.
• use_project — set the current app for subsequent tools ("switch to ").

Revenue & cost

• get_revenue — MRR, paying-customer count, per-rail split (Stripe/Apple/Google), optional daily trend.
• get_customers — the customer census: the paying-customer count (plus total / active where instrumented). The canonical "how many customers does this app have" answer.
• get_read_cost — database read-cost split into per-user reads vs un-attributed overhead, by operation.

The moat — error × identity × revenue

• get_error_impact — for one error: how many distinct users hit it, and how many are paying (counts).
• get_error_affected_users — the named paying users an error hit, each with monthly revenue and last-hit, sorted by revenue at risk. Returns your own identifiers only — never emails/names you didn't provide.
• get_customer — one customer across every layer Crossdeck joins by identity: what they pay × active entitlements × read-cost.

Analytics

• get_host_analytics — page views + unique visitors for a host/subdomain you own (optional daily series).
• get_host_top_pages — top pages or referrers for a host you own.

Rendered UI (MCP Apps)

• draw_user_growth — an interactive line chart of visitors + page views over time.
• open_moat_dashboard — a customer's cross-layer dashboard (pay × entitlements × read-cost) rendered as cards.

Security

Prism is built on Crossdeck's bank-grade outbound-read spine. See SECURITY.md for the full model. In short:

• OAuth 2.1 + PKCE (S256), dynamic client registration, short-lived (1h) workspace credentials with refresh-token rotation and reuse detection. No long-lived secret leaves your dashboard.
• Read-only. Every tool is readOnlyHint: true — no writes, no config mutations.
• Scoped & fail-closed. A token resolves to one owner and their own projects; a missing/invalid token returns an actionable 401, never silent data.
• Your data only. Prism reads your own project's data, scoped by your token. It does not read your conversation history or local files, and returns your own identifiers — never PII you didn't provide.

Known limitations

• Read-only in v1 — writes/config are a later, separately-gated stage.
• Per-host analytics require the host to be a verified origin of your project.
• Rendered charts require an MCP-Apps-capable host; others get a text summary.
• Reporting reflects data from when Crossdeck was connected (historical backfill is separate).

Links

• npm: https://www.npmjs.com/package/@cross-deck/ai
• API reference: https://cross-deck.com/docs/reporting-api/
• Security spine: https://cross-deck.com/docs/outbound-api-security/
• Privacy: https://cross-deck.com/legal/privacy/
• Support: [e-mail]

License

MIT © Crossdeck