Arcate MCP Server

Your AI agent doesn't know who your customers are. It doesn't know which ones pay the most, which ones are churning, or which problems have been reported twelve times. Every session starts from scratch — and the output shows it.

Arcate fixes that. One config block, and your agent gets structured, revenue-weighted customer evidence it can query mid-conversation.

You: "What should we build next?"

Agent (via rank_initiatives):

  1. Remove the Prompt Ceiling for Agencies — Score: 21K (High Leverage)
     $487K ARR across 32 accounts, 47 signals
  2. Tell Users What To Do Next — Score: 1.0K (Medium Leverage)
     $332K ARR across 37 accounts, 56 signals
  3. Launch a Public REST API — Score: 3.3K (High Risk)
     $221K ARR but thin evidence — needs more validation

That's not a mockup. That's the actual output from a live workspace with real customer data, scored by impact.

Setup (2 minutes)

1. Generate an API Key

Log in → Settings → Integrations → Generate API Key. Copy the key — shown only once.

2. Add to your MCP client

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "arcate": {
      "serverUrl": "https://mcp.arcate.io",
      "headers": {
        "Authorization": "Bearer arc_YOUR_KEY_HERE"
      }
    }
  }
}

Cursor → Settings → MCP → Add Server → Type: HTTP → URL: https://mcp.arcate.io
Header: Authorization: Bearer arc_YOUR_KEY_HERE

3. Test it

"What are my highest-impact initiatives? Which ones have enough evidence to commit resources?"

No installation. No npm. It's a remote HTTP server — configure the URL and go.

Requires an active Evidence subscription (€129/mo). API keys are generated in /settings/integrations inside your Arcate workspace.

What Your Agent Gets

Resources (live data streams)

Read Tools

Write Tools

What the Agent Does With It

"What should we build next?"

Your agent calls rank_initiatives. Every initiative comes back scored by a Fermi leverage model that weights:

• Revenue at risk — log-scaled ARR across all linked customer accounts

• Signal strength — type-weighted (deal-loss > problem > friction > mention), sqrt-dampened

• Evidence breadth — confirmation bonus for signals from multiple independent accounts

• Risk detection — high ARR with thin evidence gets flagged as "High Risk"

The labels tell you what to do:

• High Leverage — evidence supports committing resources

• Medium Leverage — promising, worth deeper investigation

• High Risk — significant revenue at stake but insufficient evidence — validate before building

• Low Confidence / Negligible — weak signal, deprioritize

"Log this customer feedback"

After a call, paste your notes. The agent:

1. Resolves the customer via search_customers

2. Creates structured signals via batch_create_signals

3. Links them to relevant initiatives via link_to_initiative

4. The impact scores update automatically

"Triage my inbox"

The agent calls search_signals with unlinked_only: true, groups by severity and type, and suggests which initiatives each signal belongs to. If a cluster has no matching initiative, it creates one.

Guided Prompts

These appear as clickable flows in Claude's prompt picker and Cursor's slash commands:

enrich_initiative Schema

target_outcome — defines the expected outcome:

{
  "target_description": "Reduce prompt-ceiling churn by 60% within 90 days",
  "metric": "churn",
  "validation_window_days": 90
}

health_metrics — key-value pairs where values must be numeric:

{
  "Adoption Rate": 0,
  "Retention Rate": 85,
  "Time to Value": { "value": 14, "type": "duration" },
  "Expansion Rate": { "value": 1.5, "type": "ratio" }
}

Valid metric types: percentage (default for plain numbers), ratio, currency, duration, number.

⚠️ Passing string values for health metrics will return a validation error.

Bootstrapping Your Roadmap

Already have a large backlog of signals? Use the Roadmap Bootstrap Prompt to turn your entire signal corpus into a structured, prioritized roadmap in a single session — no manual triage required.

→ Roadmap Bootstrap Prompt

Security

• Keys are stored as SHA-256 hashes. The plaintext is shown only once and never stored.

• Every request is re-validated against billing_status and use_mcp capability.

• All queries are hard-scoped to your organization_id. Cross-tenant access is impossible.

• MCP-created signals are tagged with ingestion_source: mcp for audit filtering.

Tech Spec

Architecture

Deployed as a Supabase Edge Function implementing JSON-RPC 2.0 over HTTP. A GET request to the server URL returns a human-readable info card — no MCP client needed to inspect it.

Source: src/ — TypeScript reference implementation
Deployment: Supabase Edge Functions (Deno)
Current version: v0.10.0

Database Setup

Apply the migration in supabase/migrations/add_mcp_tables.sql to bootstrap the api_keys table.