Understanding Seer MCP¶

Audience: Builders, Leadership
Type: Explanation (Understanding-Oriented)

What is Seer MCP?¶

Seer MCP is Seer's centralized Model Context Protocol platform - a cloud-hosted service that provides AI tools, data access, and workflow templates to any LLM front-end (Claude, Open WebUI, NinjaCat, etc.).

Think of it as the execution layer that turns AI "thinking" into reliable "doing."

The Core Insight¶

"Use the pen for thinking; use the calculator for certainty."

LLMs are excellent at reasoning, synthesis, and creative work. They're unreliable at arithmetic, data retrieval, and API calls. Seer MCP handles the "calculator" work so LLMs can focus on the "pen" work.

Task Type	Who Handles It	Why
Reasoning, drafting, synthesis	LLM	Creative, context-dependent
Data retrieval, API calls	Seer MCP	Must be deterministic
Calculations, scoring	Seer MCP	Must be repeatable
Workflow orchestration	LLM + MCP Prompts	Guided creativity

The Building Analogy¶

Think of it like constructing a building:

Seer MCP is structural infrastructure. It's the raw material and load-bearing skeleton - steel, concrete, protocols. It defines what can be built safely and repeatedly, but doesn't tell you what gets built.

agents-infra is the inhabitable layer. It's the floorplan, rooms, and interiors that make the building usable day to day. This is where decisions happen, where tradeoffs get made for real constraints, where usefulness appears.

One enables possibility; the other determines whether people actually want to work there.

Without Seer MCP, you'd be building without a foundation. Without agents-infra, you'd have a skeleton nobody can occupy.

Where Seer MCP Fits¶

┌─────────────────────────────────────────┐
│  What you're doing                       │
│  (Division plugins: SEO, Analytics, Ops) │
│  Commands like /content-audit            │
└─────────────────────────────────────────┘
                    │
┌─────────────────────────────────────────┐
│  How you behave                          │
│  (Skills: writing-standards, seo-methods)│
│  Automatic patterns and standards        │
└─────────────────────────────────────────┘
                    │
┌─────────────────────────────────────────┐
│  How you touch reality  ← SEER MCP       │
│  (Tools, Resources, Prompts)             │
│  Deterministic data and execution        │
└─────────────────────────────────────────┘
                    │
┌─────────────────────────────────────────┐
│  External world                          │
│  (BigQuery, DataForSEO, Wrike, etc.)     │
└─────────────────────────────────────────┘

Key insight: Your division plugins (SEO, Analytics, etc.) define what to accomplish. Skills define how to behave. Seer MCP provides reliable access to external data and systems.

Mapping the Analogy¶

Building	Technology	Purpose
Foundation & skeleton	Seer MCP	Protocols, contracts, reliable execution
Floorplan & rooms	agents-infra plugins	Workflows, commands, deliverables
Interior design	Skills	Standards, voice, methodology
Utilities (power, water)	MCP Tools	Data access, API calls
Building codes	Anthropic MCP spec	Interoperability standards
Occupants	Practitioners	People who actually use the building
Architects	Builders	People who design and construct

Why Centralized?¶

The Problem with Distributed MCP¶

Without Seer MCP, each employee would need:

Local Python environment
API credentials configured
MCP server running on their machine
Manual updates when tools change

That's ~200 people maintaining local infrastructure.

The Centralized Solution¶

With Seer MCP:

Zero local setup - Connect to cloud service
Single credential store - OAuth handles permissions
Automatic updates - Deploy once, everyone gets it
Centralized logging - Audit trail for all tool usage
Role-based access - RBAC at the tool level

What Seer MCP Provides¶

1. Tools (Deterministic Execution)¶

Tools guarantee the same output for the same input. Use them when correctness matters.

Examples:

get_search_volume(keyword, location) - Exact search volume data
query_bigquery(sql) - Database queries
resolve_location_code(location) - API code lookups

When to use: Any time you need real data, not LLM estimation.

2. Resources (Data Access)¶

Resources provide read access to data sources, respecting user permissions.

Examples:

bigquery://tables - List accessible tables
presentations://search?q=topic - Search presentation content

When to use: Browsing or retrieving structured data.

3. Prompts (Workflow Templates)¶

Prompts are long-form instructions that teach the LLM how to complete complex tasks.

Examples:

ai_search_volume_analysis - Complete workflow for volume analysis
content_audit_workflow - Step-by-step audit process

When to use: Complex multi-step workflows that need consistent execution.

4. Long-Running Jobs¶

Some tasks are too slow or expensive for synchronous execution.

Examples:

Bulk keyword research (thousands of keywords)
Cross-dataset correlation analysis
Historical data exports

Output: Delivered via email or stored artifact.

Build Once, Consume Anywhere¶

A core principle of our architecture: agents are built in one place but consumed through multiple layers.

┌─────────────────────────────────────────────────────┐
│  WHERE AGENTS ARE BUILT                              │
│  GitHub repo → Claude Code (plugins, skills, etc.)   │
└─────────────────────────────────────────────────────┘
                        │
                   Seer MCP
                (connective tissue)
                        │
        ┌───────────────┼───────────────┐
        ▼               ▼               ▼
   ┌──────────┐   ┌───────────┐   ┌───────────┐
   │ NinjaCat │   │ Claude.ai │   │ Claude    │
   │ (+ data) │   │ (general) │   │ Code      │
   └──────────┘   └───────────┘   └───────────┘
   Practitioners   Practitioners    Builders
   + client data   without data     + full tooling

Why this matters:

Builders work in Claude Code with full plugin ecosystem, version control, and automatic updates
Practitioners use whatever consumption layer fits their workflow
Seer MCP is the connective tissue that makes agents accessible everywhere

Builders¶

Who: Engineers, agent developers, power users

Where: Claude Code, OpenCode, IDE

Activities:

Create new MCP tools
Design workflows (prompts)
Build division plugins
Test and QA

Key benefit: Changes pushed to GitHub automatically propagate to all consumption layers via plugin updates.

Practitioners/Teams¶

Who: SEO strategists, analysts, account managers

Where: NinjaCat, Claude.ai, dedicated interfaces

Activities:

Run established workflows
Generate client deliverables
Access client data
Consume what builders create

This split is intentional. Builders need flexibility and version control. Practitioners need reliability and ease of use.

How Plugins Use Seer MCP¶

When you run a command like /content-audit, here's what happens:

1. User runs /content-audit command (agents-infra)
2. Command loads relevant skills (writing-standards, seo-methods)
3. LLM orchestrates the workflow
4. LLM calls Seer MCP tools for data:
   - get_serp_data(url) → SERP analysis
   - query_bigquery(sql) → Performance metrics
   - get_search_volume(keywords) → Volume data
5. LLM synthesizes results using skill standards
6. Output delivered in Seer voice (writing-standards)

The LLM does the thinking. Seer MCP provides the facts.

Graceful Degradation¶

What if Seer MCP is unavailable? Commands should handle this gracefully:

Scenario	Behavior
MCP available	Use tools for data retrieval
MCP unavailable	Prompt user for manual data input
Partial availability	Use available tools, fallback for others

Example (from command definition):

## Data Sources
1. **Primary**: Seer MCP (BigQuery, DataForSEO)
2. **Fallback**: Manual CSV upload
3. **Alternative**: GA4 export, Search Console data

Consumption Layers Compared¶

A consumption layer is where practitioners interact with an LLM to do their work. Each has different strengths:

NinjaCat¶

Aspect	Details
What it is	Chat interface built by a marketing reporting company
Unique value	Client data already connected (Search Console, GA4, Google Ads via Snowflake)
Best for	Practitioners who need to query client data while chatting
Limitation	Not designed for agent versioning or complex deterministic workflows

Why NinjaCat matters: It's the only consumption layer where client marketing data is already aggregated and available. Practitioners don't need to export CSVs or configure connections—they can ask questions about client performance directly.

Claude.ai / Claude Teams¶

Aspect	Details
What it is	Anthropic's web chat interface
Unique value	Can load skills directly; familiar chat UX
Best for	Quick skill-based work without client data needs
Limitation	Manual skill updates (no auto-sync from GitHub); no built-in data connections

How skills work here: Skills can be uploaded to Claude.ai, but updates require manually replacing the skill file. No automatic versioning.

Claude Code / OpenCode¶

Aspect	Details
What it is	Technical harness with full development tooling
Unique value	Plugin marketplace, automatic updates, version control integration
Best for	Builders; power users comfortable with terminal
Limitation	Requires technical comfort; not for everyone

Why builders use this: When you run /plugin install seo@seer-agent-core, you get automatic updates. Push a fix to GitHub → everyone on Claude Code gets it immediately.

Choosing the Right Layer¶

If you need...	Use...
Client data + chat in one place	NinjaCat
Quick skill-based task, no data	Claude.ai
Building/testing agents	Claude Code
Full automation + version control	Claude Code

How Seer MCP Connects Them¶

Seer MCP is the connective tissue that makes agents work across all consumption layers:

┌─────────────────────────────────────────────────────────────┐
│                        Seer MCP                              │
│  Tools: get_search_volume(), query_bigquery(), etc.          │
│  Resources: bigquery://tables, presentations://search        │
│  Prompts: content_audit_workflow, competitor_analysis        │
└─────────────────────────────────────────────────────────────┘
        ▲               ▲               ▲
        │               │               │
   NinjaCat         Claude.ai       Claude Code
   (via custom      (via MCP        (via plugin
    action)          config)         install)

Result: The same agent built in GitHub works in NinjaCat, Claude.ai, or Claude Code. Practitioners choose their preferred interface; the underlying capabilities are identical.

NinjaCat: Learnings and Best Fit¶

Based on our experience building and deploying agents, here's what we've learned about NinjaCat's strengths and limitations:

NinjaCat is Good For¶

Use Case	Why
Non-signals marketing data	Already connected to Search Console, GA4, Google Ads, etc.
Team access to LLM chat	Familiar interface, no technical setup required
Low-tech prototyping	Quick experiments without code
Exploratory data analysis	Chat with your client data directly

NinjaCat is Not Ready For¶

Use Case	Why	Alternative
Building agents at scale	No version control, no plugin ecosystem	Build in Claude Code, consume via NC
Deterministic workflows	Chat-based, not programmable	Use Seer MCP tools
Versioning and rollback	No GitHub-style history	Build in GitHub repo
Complex multi-step automation	Limited orchestration	Use Claude Code + MCP

The Hybrid Approach¶

Build outside NinjaCat. Consume through NinjaCat (when data is needed).

This gives us:

✅ Version control and CI/CD for agent code (GitHub)
✅ Fast iteration with full tooling (Claude Code)
✅ Client data access for practitioners (NinjaCat)
✅ Flexibility to use other consumption layers (Claude.ai, etc.)

Example flow:

Builder creates /content-audit command in GitHub
Command uses Seer MCP for data retrieval
Practitioner opens NinjaCat, connects to Seer MCP
Practitioner runs the audit with client data already available
Builder pushes fix → Seer MCP updates → Practitioner gets fix automatically

Planned: Snowflake Integration (NinjaCat Data)¶

A key upcoming integration will bring NinjaCat's client data to all consumption layers via Snowflake.

The Problem¶

Currently, client marketing data (Search Console, GA4, Google Ads) is only accessible through NinjaCat's chat interface. Practitioners using Claude.ai or Claude Code can't query this data directly.

The Solution¶

Integrate the official Snowflake MCP server (Snowflake-Labs/mcp) into Seer MCP's authorization layer.

┌─────────────────────────────────────────────────────────────────┐
│                         Seer MCP                                 │
│                    (Authorization Layer)                         │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐  │
│  │ BigQuery    │  │ DataForSEO  │  │ Snowflake               │  │
│  │ (internal)  │  │ (SERP data) │  │ (NinjaCat client data)  │  │
│  │             │  │             │  │                         │  │
│  │             │  │             │  │ snowflake-labs-mcp      │  │
│  └─────────────┘  └─────────────┘  └─────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘

Why the Official Snowflake MCP¶

Attribute	Value
Repo	Snowflake-Labs/mcp
Maintainer	Snowflake Labs (official)
License	Apache 2.0
Stars	200+
Install	`uvx snowflake-labs-mcp`

Key features:

SQL execution with granular permission controls
Object management (list databases, schemas, tables)
Cortex AI integration (semantic queries, RAG)
Multiple auth methods (OAuth, key pair, SSO)
Docker/container deployment support

What This Enables¶

Before	After
Client data only in NinjaCat	Client data in ANY consumption layer
Must use NC for data queries	Query from Claude.ai or Claude Code
Data siloed from agents	Unified data access via Seer MCP

Planned Configuration¶

The Snowflake MCP will be wrapped by Seer MCP's authorization layer, with read-only access for most users:

# Planned: Snowflake service configuration
sql_statement_permissions:
  - Select: True      # Read queries allowed
  - Describe: True    # Schema inspection allowed
  - Insert: False     # Write operations disabled
  - Update: False
  - Delete: False
  - Drop: False

Status¶

Phase	Status
Research & selection	✅ Complete (official Snowflake-Labs/mcp selected)
Technical specification	🟡 In progress (ENG team)
Integration with Seer MCP auth	⏳ Planned
Testing with NinjaCat data share	⏳ Planned
Production deployment	⏳ Planned

Owner: Engineering team (Seer MCP)

Security Model¶

Authentication¶

OAuth-based - Single sign-on with Seer credentials
Token refresh - Transparent session management
No local credentials - Users don't handle API keys

Authorization (RBAC)¶

Role-based - user, alpha_tester, admin
Tool-level - Each tool declares required roles
Inherited permissions - BigQuery access follows existing IAM

Audit¶

Centralized logging - All tool calls logged
Cost tracking - API usage monitoring
Access audit - Who used what, when

Key Takeaways¶

Build once, consume anywhere - Agents built in GitHub work across NinjaCat, Claude.ai, and Claude Code
Seer MCP is the connective tissue - It makes agents accessible from any consumption layer
Seer MCP is the "calculator" - Use it for anything that needs to be deterministic
LLMs are the "pen" - Use them for reasoning, synthesis, creativity
NinjaCat's unique value is data - It's where client marketing data already lives
Builders create, practitioners consume - Build in Claude Code, use anywhere

Seer MCP Technical Specification - Full technical details
Plugin Architecture - How plugins are structured
MCP Servers Configuration - Local setup

Understanding Seer MCP¶

What is Seer MCP?¶

The Core Insight¶

The Building Analogy¶

Where Seer MCP Fits¶

Mapping the Analogy¶

Why Centralized?¶

The Problem with Distributed MCP¶

The Centralized Solution¶

What Seer MCP Provides¶

1. Tools (Deterministic Execution)¶

2. Resources (Data Access)¶

3. Prompts (Workflow Templates)¶

4. Long-Running Jobs¶

Build Once, Consume Anywhere¶

Builders¶

Practitioners/Teams¶

How Plugins Use Seer MCP¶

Graceful Degradation¶

Consumption Layers Compared¶

NinjaCat¶

Claude.ai / Claude Teams¶

Claude Code / OpenCode¶

Choosing the Right Layer¶

How Seer MCP Connects Them¶

NinjaCat: Learnings and Best Fit¶

NinjaCat is Good For¶

NinjaCat is Not Ready For¶

The Hybrid Approach¶

Planned: Snowflake Integration (NinjaCat Data)¶

The Problem¶

The Solution¶

Why the Official Snowflake MCP¶

What This Enables¶

Planned Configuration¶

Status¶

Security Model¶

Authentication¶

Authorization (RBAC)¶

Audit¶

Key Takeaways¶

Related Documentation¶