MCP Everywhere

Claude is brilliant. But Claude is blind.

Ask Claude about your production errors? It doesn't know. Ask about your API performance? No idea. Ask about your log patterns? Can't see them.

Unless you give Claude eyes.

MCP: Model Context Protocol

MCP is Anthropic's protocol for connecting AI to tools. It's simple:

1. A server exposes tools (functions Claude can call)
2. Claude discovers available tools
3. Claude calls tools to get data or take actions
4. Results flow back to Claude

Every Brainz Lab product is an MCP server.

What this means

Ask Claude: "Are there any errors in production right now?"

Claude doesn't guess. Claude calls reflex_list_errors:

{
  "tool": "reflex_list_errors",
  "parameters": {
    "status": "unresolved",
    "since": "1h"
  }
}

Reflex returns real data. Claude analyzes it. You get an answer grounded in reality.

Every product, every tool

Recall (logs):
- recall_query — Search logs with our query DSL
- recall_tail — Stream recent logs
- recall_stats — Log volume and patterns

Reflex (errors):
- reflex_list_errors — Get error list
- reflex_get_error — Error details with stack trace
- reflex_resolve_error — Mark as resolved

Pulse (traces):
- pulse_list_traces — Recent traces
- pulse_get_trace — Full trace with spans
- pulse_slow_endpoints — Performance issues

Signal (alerts):
- signal_list_incidents — Active incidents
- signal_acknowledge — Ack an incident
- signal_get_status — Overall system health

Vault (secrets):
- vault_get_secret — Retrieve a secret
- vault_list_secrets — Available secrets
- vault_generate_otp — Get current TOTP code

Flux (metrics):
- flux_query — Query custom metrics
- flux_recent_events — Recent events
- flux_anomalies — Detected anomalies

Dendrite (docs):
- dendrite_search — Semantic code search
- dendrite_ask — Ask about codebase
- kb_search — Search knowledge bases

Synapse (tasks):
- synapse_list_tasks — Current tasks
- synapse_create_task — Create a task
- synapse_ask — Consult an AI executive

The compound effect

Individual tools are useful. Combined tools are powerful.

"Why is checkout slow today?"

Claude orchestrates:

1. pulse_slow_endpoints — Find slow endpoints
2. pulse_get_trace — Get detailed trace
3. recall_query — Find related logs
4. reflex_list_errors — Check for errors
5. Synthesizes findings

One question. Multiple data sources. Complete answer.

How it works

Each product exposes an MCP endpoint:

GET  /mcp/tools           # List available tools
POST /mcp/tools/:name     # Call a tool
POST /mcp/rpc             # JSON-RPC 2.0

Claude Code connects to these endpoints. Discovers tools. Calls them as needed.

// Claude Code configuration
{
  "mcpServers": {
    "recall": { "url": "http://recall.localhost/mcp" },
    "reflex": { "url": "http://reflex.localhost/mcp" },
    "pulse": { "url": "http://pulse.localhost/mcp" }
  }
}

All products available. All tools accessible.

Authentication flows through

MCP requests include your API key. Tools respect permissions.

Can't access an organization you don't belong to. Can't query projects you don't have access to. Same security model as the REST API.

The developer experience

You're debugging an issue. You have Claude Code open.

Instead of:
1. Open Reflex dashboard
2. Search for errors
3. Copy error details
4. Open Pulse dashboard
5. Find related traces
6. Correlate manually

You just ask: "What's causing the 500 errors on /api/orders?"

Claude does the legwork. You get the answer.

Building on MCP

We expose MCP. You can too.

Hive Core includes MCP server utilities:

class MyService
  include HiveCore::MCP::Server

  mcp_tool :my_custom_tool do |params|
    # Your implementation
    { result: "data" }
  end
end

Extend the ecosystem. Give Claude more capabilities.

The vision

AI assistants should have access to everything developers have access to.

Logs, errors, traces, metrics, docs, tasks—all queryable by AI.

Not through copy-paste. Through direct integration.

MCP makes this possible. Brainz Lab makes it real.

— Andres