Stack decision view.
Cross-category comparison is a different question. We map the stack and show what fits where — not winners and losers.
Apples and oranges
Comparing Claude Desktop and Cursor as if one is ``better'' is the wrong question — they live on different floors of the agentic stack. The matrix views handle same-floor comparisons. This page handles cross-floor: which model with which host, which framework on which protocol.
Where each compare category sits
| Layer | Question | Compare category | Matrix |
|---|---|---|---|
| L01 · UI / TRIGGER | What kicks off the agent? | (emerges from Hosts choice) | — |
| L02 · ORCHESTRATION | How does the loop work? | Frameworks | → FRAMEWORKS |
| L03 · MODEL | What does the reasoning? | LLMs | → LLMS |
| L04 · TOOLS | What does the agent touch? | MCP servers · Protocols · Vendors | → PROTOCOLS |
| L05 · MEMORY / CONTEXT | What does the agent remember? | (typically embedded in framework choice) | — |
| CROSS | Where does it run? | Hosts | → HOSTS |
What pairs with what
Coding agent on a developer laptop
- L03 model: Claude Sonnet 4.5 — best tool-call reliability for code work.
- Cross host: Claude Code (terminal) or Cursor (IDE) — both are first-class.
- L02 framework: usually unnecessary — the host orchestrates.
- L04 protocol: MCP for tool connection (github-mcp, filesystem-mcp).
Workflow that runs on a server, no human in the loop
- L03 model: Claude Sonnet 4.5 or GPT-5 — pick on cost vs benchmark fit for your task.
- Cross host: none — no human means no chat host. Run from a script.
- L02 framework: Pydantic AI (type-safe, light) or LangChain (if you need many integrations).
- L04 protocol: MCP, called via your framework's MCP adapter.
Local-first, data-residency-strict
- L03 model: Llama 3.3 70B (or larger Llama variant) on Ollama / LM Studio.
- Cross host: Cline (free, BYO model) — best MCP support for local models in an IDE.
- L02 framework: only if you want headless. Otherwise the host is enough.
- L04 protocol: MCP (the local model needs MCP-aware host or framework).
Multi-agent system (one agent delegates to another)
- L03 model: mix-and-match per agent's task.
- Cross host: none typically — multi-agent runs from a framework, not a chat host.
- L02 framework: LangChain (LangGraph), Microsoft Agent Framework, or CrewAI.
- L04 protocol: MCP for tools plus A2A for agent-to-agent delegation.
These are starting points
Real architectures have constraints we can't enumerate from afar — your latency budget, your team's existing skills, your compliance posture, your existing platform commitments. Treat these as starting points and adjust.
Disagree with one of these stack decisions? Open an issue with your context — we'll add common variations as the ``stack decisions'' bench grows.
Suggest a stack decision