Skip to main content
Glama

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.

npm version License: MIT MCP OAuth 2.1 Read-only

"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.

What Prism does → · Developer docs · Crossdeck


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.


Related MCP server: ai-analyst

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

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 secret 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_sk_live_…" }
    }
  }
}

Use a secret key (cd_sk_) from your dashboard → API keys — 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:

  1. 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).

  2. 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).


License

MIT © Crossdeck

Available Tools

11 tools
draw_user_growthDraw user growth over timeA
Read-only
Inspect

Render an interactive line chart of unique visitors and page views over time for a host you own, and return the period totals as text. Returns a rendered chart in MCP-Apps-capable hosts (a text summary elsewhere) plus the totals. Use when asked to chart, graph, draw, or visualize growth or traffic for a subdomain. The host must be a verified origin of the project, or the request is rejected.

Inputs: host (required), days, project.

get_customerGet a customer's full pictureA
Read-only
Inspect

Cross-match one customer across every layer Crossdeck joins by identity. Returns what they pay (monthly cents), their active entitlement count, and their database read-cost — in one view. Identify the person by ANY ONE of the identifiers below; they all resolve to the same canonical customer. Use for 'how much does this user pay and what do they cost us?'. If no customer resolves, returns a no-match result, not an error.

Inputs: userId, project, customerId, anonymousId, +3 more.

get_error_affected_usersGet who an error hit and what they payA
Read-only
Inspect

For one error (by fingerprint), get the PAYING users it actually hit — the row-level moat answer: which paying customer broke, and how much are they worth. Returns a list sorted by revenue at risk (highest first), each row with your own user id, monthly revenue in cents, and when they last hit the error. Joins the error layer to identity and revenue. Returns only your own identifiers — never emails or names you didn't provide. Use when you need the actual customers, not just the counts from get_error_impact.

Inputs: fingerprint (required), limit, project.

get_error_impactGet an error's impactA
Read-only
Inspect

For one error (by fingerprint/issue id), get its blast radius joined to identity. Returns distinct users affected, how many of those are PAYING customers, and the error's type, status, occurrence count, and first/last-seen timestamps — counts only. Use to size an error's impact. For the NAMED paying users behind the counts (and their revenue at risk), call get_error_affected_users instead.

Inputs: fingerprint (required), project.

get_host_analyticsGet analytics for a hostA
Read-only
Inspect

Get headless web analytics for one host/subdomain the app owns (e.g. a tenant's subdomain). Returns page views and unique visitors with totals; granularity='day' adds a daily series. The host MUST be a verified origin of the project, or the request is rejected (403). Use for per-tenant or per-subdomain traffic questions.

Inputs: host (required), days, project, granularity.

get_host_top_pagesGet top pages or referrers for a hostA
Read-only
Inspect

Get the top pages or top referrers for one host the app owns, ranked and paginated. Returns rows of { key, views } highest-first, where key is a page path (default) or a referrer. Set dimension='top_referrers' for traffic sources. The host must belong to the selected app. Use for 'most-viewed pages on this subdomain' or 'where its traffic comes from'.

Inputs: host (required), days, limit, project, +1 more.

get_read_costGet database read-costA
Read-only
Inspect

Get an app's database read-cost over the last days days, split into per-user reads vs un-attributed overhead, plus a breakdown by operation. Per-user attribution works because Crossdeck joins read-cost to the SDK's identity, so you can see which operations drive the bill. Returns the per-user-vs-overhead totals and the by-operation breakdown. Use for 'what's driving our database reads' or 'which operation costs the most'.

Inputs: days, project.

get_revenueGet revenueA
Read-only
Inspect

Get an app's recurring revenue from its maintained revenue ledger (a point-read, never a scan). Returns MRR in cents, the paying-customer count, and the per-rail split across Stripe, Apple, and Google; with granularity='day' it adds a daily time series over the window. Use for 'what's our MRR', 'how many paying customers', or revenue-trend questions. An app with no revenue yet returns zeros, not an error.

Inputs: days, project, granularity.

list_projectsList your appsA
Read-only
Inspect

List every app (project) in your Crossdeck portfolio that this connector can read — the same set your Pulse dashboard shows. Returns an array of { id, name } per app. Call this FIRST when a workspace key (cd_wk_) is connected, then pass an id to other tools or set a default with use_project. With a single-app key it returns just that one app. Read-only; takes no arguments.

Inputs: no named parameters.

open_moat_dashboardOpen the cross-layer dashboardA
Read-only
Inspect

Render a customer's cross-layer dashboard — what they pay (monthly), their active entitlements, and their database read-cost, joined by identity — and return the same as a text summary. Returns a rendered dashboard in MCP-Apps-capable hosts (a text summary elsewhere). Identify the customer by ANY ONE of the identifiers below; they all resolve to the same canonical customer. Use when asked to show, open, or visualize a customer's full picture. Returns a no-match result if no customer resolves.

Inputs: userId, project, customerId, anonymousId, +3 more.

use_projectSelect the current appA
Read-only
Inspect

Set the current app for this conversation so later tools default to it without repeating project. Behaves as session state: it persists until you call this again and applies to every subsequent tool call. Returns { currentProject } — the project id now in effect. Use when the user says 'switch to ' or names an app to focus on; call list_projects first for valid ids. No effect with a single-app key (the project is fixed).

Inputs: project (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 11 tools

    Disambiguation4/5

    Most tools target distinct resources (projects, customers, errors, hosts, revenue, costs). However, get_customer and open_moat_dashboard both return overlapping customer info (payment, entitlements, read-cost) with different output formats, which could cause selection ambiguity.

    Naming Consistency4/5

    Tools follow a consistent imperative_underscore pattern with verbs like get_, list_, draw_, open_, and use_. The naming is clear and predictable, though the verb choice varies appropriately across actions.

    Tool Count5/5

    With 11 tools covering project management, customer insights, error analysis, host analytics, and revenue/costs, the set is well-scoped for a product analytics platform. Each tool serves a distinct purpose without redundancy.

    Completeness4/5

    The core analytics needs are covered: customer overview, error impact, traffic analysis, revenue tracking, and cost breakdown. Minor gaps like user-level event logs or data export are acceptable given the server's focus on cross-layer analytics.

    Maintenance

    ActivityStale
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI-powered analytics from Stripe, PayPal, and Google Analytics 4 (BigQuery) data sources with built-in guardrails and automated workflows for financial and web performance insights.
      ISC
    • A
      license
      Not graded
      quality
      F
      maintenance
      Enables natural language data queries with statistical rigor and a semantic layer, supporting 23+ analytical tools for product analytics.
      1
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables natural language querying of databases with multi-turn conversations, auto-generated charts, and proactive monitoring via scheduled queries and alerts.
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables asking product analytics questions in English and receiving answers as real SQL along with the resulting numbers, through read-only MCP tools for funnels, retention, trends, paths, and event counts.
      MIT