Skip to content

MCP Servers

Nova AI uses 8 in-process MCP servers for 69x faster tool execution.

Why In-Process?

Traditional MCP uses stdio (subprocess communication):

Agent → spawn process → JSON-RPC → tool → JSON-RPC → result
         ~100ms overhead per call

Nova AI uses SDK MCP (in-process):

Agent → direct function call → result
         ~1.5ms overhead per call

Result: 69x faster tool execution.

Available Servers

Knowledge Base (sdk_mcp_kb_server)

5.4GB FAISS vector database for semantic search.

# Search for patterns
results = await search_knowledge_base(
    query="JWT authentication best practices",
    top_k=5
)

Use cases: - Find coding standards - Look up security patterns - Search API conventions

GitHub (sdk_mcp_github_server)

GitHub API operations without leaving the agent.

# Create a PR
await create_pull_request(
    title="Add auth feature",
    body="Implements JWT authentication",
    base="main"
)

Use cases: - Create/update PRs - Manage issues - Read commit history

Memory (sdk_mcp_memory_server)

Persistent storage across sessions.

# Store context
await memory_store("user_preferences", {"theme": "dark"})

# Retrieve later
prefs = await memory_get("user_preferences")

Use cases: - Remember user preferences - Cache expensive computations - Track session state

Blarify (sdk_mcp_blarify_server)

Code analysis and dependency graphing.

# Check for circular dependencies
issues = await check_circular_deps("src/")

# Analyze module dependencies
graph = await analyze_dependencies("src/auth/")

Use cases: - Detect circular imports - Find dead code - Map dependencies

Composio (composio_app_tools)

100+ app integrations.

Supported apps: - Slack, Discord - Jira, Linear - Notion, Confluence - Gmail, Calendar

Context7

Live framework documentation.

# Get latest React docs
docs = await get_docs("react", "hooks")

Use cases: - Up-to-date API docs - Framework examples - Migration guides

AWS Pricing

Cost estimation for deployments.

# Estimate monthly cost
cost = await estimate_cost({
    "ec2": "t3.medium",
    "rds": "db.t3.micro"
})

CI Triage

GitHub Actions failure analysis.

# Analyze failed workflow
analysis = await triage_failure(
    repo="owner/repo",
    run_id=12345
)

Performance Comparison

Operation stdio MCP SDK MCP
KB search 150ms 2ms
GitHub API 200ms 3ms
Code analysis 500ms 8ms

Configuration

MCP servers are registered in .claude/settings.json:

{
  "mcpServers": {
    "kb": {
      "type": "sdk",
      "module": "src.orchestrator.tools.sdk_mcp_kb_server"
    }
  }
}

What's Next?