Concept card

LangChain / LlamaIndex / Mosaic AI vs MCP — Framework vs Protocol

As of 2026-07-24T14:00:00Z

Edited by Cédric-Olivier Mâry, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is LangChain / LlamaIndex / Mosaic AI vs MCP — Framework vs Protocol?

Confusing an agent framework (LangChain, LlamaIndex, Mosaic AI) with a protocol (MCP) is the most common 2026 agent-architecture mistake — the framework choice is reversible in weeks, the tool-catalogue protocol choice is not.

What This Comparison Actually Is

LangChain (originating in Python and JavaScript in October 2022), LlamaIndex (originally GPT-Index, launched January 2023), and Databricks Mosaic AI Agent Framework (acquired by Databricks in June 2024, generally available on the Databricks platform from October 2024) are agent application frameworks: opinionated software development kits that bundle tool-calling, prompt orchestration, memory, retrieval, evaluation, and tracing into one developer surface. The Model Context Protocol (MCP) is something categorically different — a wire protocol, a contract between an agent and a catalogue of tools, with no opinion at all about how the agent itself is built. Treating these as competing choices is the single most common architecture mistake in agent projects, because the question "framework or protocol?" is not actually a choice — you need both, and they operate at different layers of the stack.

How the Two Layers Fit Together

Why it matters

Switching LangChain for LlamaIndex costs roughly 2-6 weeks on a mid-size project; migrating 20 tools between protocols breaks every consuming agent at once and typically freezes new development for 3-6 months.
A LangChain agent can be an MCP client, and a LangChain-built tool can be exposed as an MCP server so non-LangChain agents can call it too.
Hard-coding a Datasphere tool against LangChain's BaseTool interface locks it out of Claude Code, Joule Agent, or Copilot Studio — exposing it as MCP instead lets every host call it.

Key points

Categories — LangChain / LlamaIndex / Mosaic AI are agent frameworks (how you build the agent); MCP is a wire protocol (how the agent talks to tools).
Complementary not competitive — a LangChain agent can be an MCP client; a LangChain tool can be exposed as an MCP server, callable from Joule or any other MCP host.
Adoption signals — LlamaIndex MCP client 2025-01; Mosaic AI Gateway registers MCP servers as managed catalogues; LangChain MCP adapter shipped community-side 2025-02.
Frameworks are reversible — switch LangChain to LlamaIndex with rewrite cost bounded by team productivity (2-6 weeks mid-size project); protocols are not — switching tool protocols breaks every consuming agent simultaneously.
SAP analytics rule — build the tool catalogue as MCP first; pick the framework second on developer ergonomics + evaluation tooling + team skills.
Lock-in asymmetry — BaseTool lock-in traps tools inside one framework; MCP server exposure lets the same tool be consumed by Joule, Copilot Studio, Claude Code, and LangChain agents without modification.
Governance lives at the protocol layer — access control, versioning, cost metering, and audit logging belong in the MCP server or the AI Agent Hub above it, not inside the framework's tool implementation.
Evaluation tooling is a framework differentiator — LangSmith (LangChain), LlamaCloud eval (LlamaIndex), and Mosaic AI evaluation (Databricks MLflow) differ materially; pick the framework whose eval harness matches the team's debugging workflow.
Multi-framework estates are common — a single SAP customer may run Joule (SAP-native), a LangChain agent (custom), and Copilot Studio (Microsoft); MCP is the only practical shared tool layer across all three.
Scaling criterion — do not scale a framework-locked tool catalogue beyond 5 tools or 3 agent hosts without migrating to MCP; the combinatorial rewrite cost grows as O(tools × hosts).

Terms used on this page

Agent framework: Opinionated SDK bundling tool-calling, prompt orchestration, memory, retrieval, evaluation, and tracing for agent application development; LangChain, LlamaIndex, Mosaic AI are exemplars.
Wire protocol: Contract between two software components specifying message format, discovery, and exchange rules; MCP is the agent-to-tool wire protocol; A2A (Agent-to-Agent) is the peer-agent one.
MCP (Model Context Protocol): Open protocol published by Anthropic (2024-11) that standardises how agents discover and invoke tools; framework-agnostic — any agent built in any framework can call any MCP server.
Mosaic AI Gateway: Databricks-managed surface (2024-10 GA) for routing agent traffic to multiple models and tool catalogues; registers MCP servers as first-class managed catalogues with governed access.
BaseTool lock-in: Anti-pattern of implementing a tool directly against a framework's class hierarchy (e.g. LangChain BaseTool) rather than as an MCP server; creates an O(tools × hosts) rewrite cost when adding new agent hosts.
Server-prefixed tool naming (MCP): MCP convention where a client or proxy aggregating tools from multiple MCP servers prefixes each tool name with a server identifier (e.g. datasphere_query vs crm_query) to avoid naming collisions when two servers expose a similarly named tool; a disambiguation convention, not a shared-deployment or governance capability.
LangSmith: LangChain's managed evaluation and observability platform; provides prompt regression testing, trace replay, and dataset-level eval — the primary differentiator for teams choosing LangChain over LlamaIndex.
A2A (Agent-to-Agent protocol): The peer-agent wire protocol (C229) that complements MCP: where MCP governs agent-to-tool calls, A2A governs calls between agents — enabling agent delegation, sub-agent orchestration, and multi-agent pipelines.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →