Foundry IQ Is Now GA Serverless Knowledge Bases, MCP Server, and Agentic Retrieval That Cuts Token Spend

The Knowledge Layer Has Become the Agent Bottleneck

Every developer building agent fleets eventually hits the same wall: the agent logic is straightforward, but the knowledge infrastructure underneath is deceptively complex. Production-ready agents demand stability, scale, data access, answer quality, security, and content ingestion — all working simultaneously.

Microsoft Foundry IQ, announced at Build 2026, directly targets this problem. It's positioned as an enterprise knowledge platform that grounds agents in your company's collective intelligence: documents, emails, meetings, operational data, and live web content. The platform now reaches general availability with several major updates.

Here's what changed and what it means for your agent architecture.

What's New in Foundry IQ: GA, Serverless, and MCP

Knowledge Bases Are Now Generally Available

The core knowledge base functionality has reached GA with full SLA coverage, compliance certifications, and stable APIs. This includes the Foundry IQ MCP server, which exposes knowledge bases as a remote MCP server compatible with Claude, ChatGPT, LangChain, and the Microsoft Agent Framework.

# Example: Connecting to Foundry IQ MCP server from Python
# 예시: Python에서 Foundry IQ MCP 서버에 연결하기

import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

async def query_foundry_iq(query: str, knowledge_base_id: str):
    """Query a Foundry IQ knowledge base via MCP"""
    server_params = StdioServerParameters(
        command="npx",
        args=["-y", "@foundry/iq-mcp-server"],
        env={
            "FOUNDRY_IQ_ENDPOINT": "https://your-service.search.windows.net",
            "FOUNDRY_IQ_API_KEY": "your-api-key",
            "FOUNDRY_IQ_KB_ID": knowledge_base_id
        }
    )
    
    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()
            
            # Perform agentic retrieval
            # 에이전트 검색 수행
            result = await session.call_tool(
                "retrieve_agentic",
                arguments={
                    "query": query,
                    "reasoning_effort": "minimal",
                    "top_k": 10
                }
            )
            return result

# Usage / 사용 예시
# result = asyncio.run(query_foundry_iq("What was Q3 revenue?", "kb-123"))

Serverless Tier: Scale to Zero Pricing

Agent workloads are bursty by nature — hundreds of steps in seconds, then idle for hours. The new Foundry IQ Serverless (Developer tier, public preview) eliminates cluster management and idle costs entirely. Billing starts September 13, 2026.

Regional pricing (CU/hr / GB-month):

• West Central US: $0.288 / $0.239
• Switzerland North: $0.343 / $0.286
• Japan East: $0.290 / $0.289

New Knowledge Sources in Preview

• Work IQ: Emails, meetings, files, Teams messages — respecting user permissions
• Fabric IQ: Query data agents and company ontologies (OneLake semantic layer)
• File Search: Direct file uploads
• Azure SQL: Structured relational data
• MCP Server: Any knowledge served over Model Context Protocol

Agentic Retrieval: 54% Better Recall, Fewer Tokens

Foundry IQ's iterative retrieval loop now batches queries more effectively, uses semantic ranking for better passage surfacing, and applies server-side token caching. The results are impressive:

• Answer quality improved up to 20% across evaluated datasets
• Recall improved up to 54% compared to single-shot RAG
• Meaningfully fewer tokens spent without quality loss

Security and Data Pipeline Updates

Several security capabilities are in preview: cross-tenant customer-managed keys (CMK) with federated identity, Purview sensitivity-label auditing, incremental SharePoint permissions sync, and APIM support. Private connectivity via Shared Private Link and Network Security Perimeter is now GA.

Data pipeline improvements include layout-aware ingestion that converts diagrams, charts, and scanned images into meaningful text — so agents are grounded in complete documents, not just raw text.

Limitations and Watch-Outs

1. Serverless is preview-only: The Developer tier is public preview. Production serverless workloads need to wait for GA pricing and SLAs.
2. Billing date is fixed: If you provision serverless resources now, delete them before September 13, 2026 to avoid charges.
3. Index limits: 30 indexes per service and 5 services per subscription per region — plan your multi-tenant architecture accordingly.
4. MCP is still evolving: While the MCP server works with Claude and ChatGPT, the broader MCP ecosystem is young. Test integration paths carefully.

Bottom Line: Build Once, Reuse Everywhere

Foundry IQ's GA marks a significant step toward standardizing the enterprise knowledge layer. The MCP server is the most interesting piece — it decouples knowledge infrastructure from any specific agent framework, which is exactly what the ecosystem needs.

Next steps for your learning path:

1. Start with the Mastering Foundry IQ cookbook to build with the MCP server
2. Experiment with the serverless tier for bursty workloads
3. Evaluate agentic retrieval quality against your own RAG baseline
4. Review the security previews (CMK, Purview labels) for your compliance requirements

For broader context on how infrastructure disruptions affect AI deployments, check out this analysis of Q1 2026 internet disruptions and their impact on cloud services. Also, if you're exploring the GPU side of AI infrastructure, our coverage of CUDA 13.3 with Tile C++ and auto-tuning compiler is worth reading.

Related resources:

• Foundry IQ documentation
• Microsoft Web IQ announcement