Comparisons
Pick the right AI tool, fast
Honest, side-by-side comparisons of the tools engineering teams are evaluating — with a clear recommendation, not a fence-sit.
AI coding tools
Claude Code vs Cursor
→Both put a capable coding agent in your workflow, but they live in different places: Claude Code is a terminal-native agent, Cursor is an AI-first editor. Most teams end up using both — the question is which becomes the default for which work.
Cursor vs GitHub Copilot
→Copilot pioneered inline completion and now ships an agent mode; Cursor rebuilt the editor around AI. The gap is narrowing, so the decision is mostly about depth of AI integration versus fitting into your existing GitHub-centric workflow.
Claude Code vs GitHub Copilot
→These solve different shapes of problem. Copilot is strongest at in-editor assistance across millions of developers; Claude Code is strongest at autonomous, terminal-driven tasks you can script and automate.
LLM routing & gateways
LiteLLM vs OpenRouter
→Both give you one interface to many models, but at different layers: LiteLLM is an open-source library/proxy you run; OpenRouter is a hosted API and marketplace. The choice hinges on control versus convenience.
LiteLLM vs Portkey
→Both are AI gateways with routing, caching, and observability. LiteLLM is open-source-first; Portkey is a commercial gateway with a managed offering. The trade is between self-hosted control and managed convenience plus enterprise features.
LLM observability & evals
LangSmith vs Langfuse
→Both do tracing, evals, and prompt management for LLM apps. LangSmith is the commercial platform from the LangChain team; Langfuse is open-source-first with a hosted option. The decision is open-source/self-host versus an integrated commercial suite.
LangSmith vs Helicone
→LangSmith is an end-to-end LLMOps platform; Helicone is a proxy-based observability and gateway tool that’s simple to drop in. The difference is depth of eval tooling versus speed of integration.
Helicone vs Langfuse
→Two popular open-source-friendly observability tools. Helicone leans proxy-first and dead simple to adopt; Langfuse leans SDK-based tracing with stronger eval and prompt features.
LLM techniques
RAG vs Fine-tuning
→A constant question: should you retrieve context at query time (RAG) or bake knowledge into the model (fine-tuning)? They solve different problems and are often combined.
Self-hosted LLMs vs API models
→Run open-weight models on your own infrastructure, or call hosted frontier APIs? The answer is driven by data sensitivity, scale economics, and how close to frontier quality you need to be.
MCP vs function calling
→These are often framed as competitors but operate at different layers. Function calling is how a model requests a tool; MCP is a standard protocol for exposing tools and data to any model or client.
How we decide, and why we take a position
A comparison that lists features in two columns and concludes that it depends is not a comparison; it is a table with an apology attached. Every page here ends with a recommendation and the conditions under which it holds, because that is the useful artefact when you are the person who has to choose by Friday. Where we pick one, we also name what the other option is genuinely better at, and those sections are written to be believed rather than as a courtesy.
The criteria are consistent across the categories. What does the thing cost you to operate, not just to license. How hard is it to get out of once you are in, which is the question teams most often skip and most often regret. How well does it behave at the scale you are actually at rather than the scale on the landing page. Who has to maintain it, and do you have that person. And what happens when it breaks at three in the morning — which is where self-hosted and managed options genuinely diverge, and where a feature table tells you nothing at all.
We do not publish another company's pricing or license terms unless we can point at where they state it themselves. Prices change within months, a rate quoted second-hand cannot be checked by you, and a claim about someone else's licensing is an assertion we would be making on their behalf. Where the commercial shape matters to the decision, we describe the shape — open source with a managed option, usage-priced, seat-priced — without attaching a number we cannot source.
None of these tools pay us and we do not run affiliate links, which is worth stating because the category norm is otherwise. What we do have is a bias from experience: we have run some of these in production for clients and not others, and where that is true the page says so. An opinion formed by operating something is more useful than a neutral summary, provided you know which one you are reading.