AI Chatbot Platforms, Coding Agents, and Model Families

A layer-by-layer comparison of the AI ecosystem stack: LLM serving, client agents, model families, protocol transport, and UI rendering. The goal is to separate what’s converged and shared across the whole space from what actually differs — ecosystem by ecosystem, and model family by model family — so the tables below can stay focused on the differences that matter.

Common ground

• The middle layers have converged: Markdown output, SSE + JSON delta streaming, Markdown → AST → React rendering, and MCP as the tool-calling standard.
• Real differences sit in model behavior, reasoning quality, context window reliability, product UX, and infrastructure strategy.
• Open weights aren’t rare: DeepSeek, Qwen, Llama, GLM, and Kimi all ship open checkpoints — DeepSeek and Qwen are the most practical picks for self-hosting and reproducibility.
• Claude Code, Codex, GitHub Copilot, and OpenCode are coding clients built on top of these platforms, not independent model stacks.
• Most internal architecture details remain proprietary; treat vendor-unconfirmed claims as estimates.

AI agent ecosystems

There’s no single winner here — the right ecosystem depends on what you’re actually optimizing for, which is why “Best for” is a column rather than a ranking.

Deep client / agent comparison

The ecosystem table above is the what; this table is the how. It digs into the internal functioning of each coding agent — the “motor” (the loop that drives the agent forward) and the “flux” (how state, context, and tool results flow through that loop). Most of this is reverse-engineered from observable behavior, so treat vendor-unconfirmed internals as estimates.

What’s shared (kept out of the table): every agent here uses MCP as its tool-calling standard (Codex is MCP-native), and nearly all run locally with permission prompts, keeping a local session transcript with tool results appended. The table below only captures what actually differs.

Reading the table. Motor is the decision loop that turns a goal into actions — most agents converge on a ReAct-style (reason → act → observe) loop, while autonomous agents swap in a goal-driven loop with memory. Flux is how context and tool results move through that loop: session transcripts, file diffs, cloud snapshots, or persistent memory stores. The columns that actually separate the field are sandboxing (cloud vs. local), multi-agent parallelism (single-threaded vs. sub-agent dispatch), and storage — where state lives on disk and whether it’s plain JSON files or a real database (SQLite). Paths and database choices are estimates from observable behavior and may vary by version.

Routers & gateways

OpenRouter, LiteLLM, Portkey, Requesty, and OmniRoute aren’t ecosystems of their own — they’re a routing layer that plugs into the clients above (VS Code Copilot Chat, Claude Code, Hermes agent, and so on), giving each one API key to reach hundreds of models across every provider instead of a separate integration per vendor. OpenRouter is the hosted default (400+ models, near-zero markup, free-tier models available); LiteLLM and OmniRoute are open-source and self-hostable (OmniRoute is local-first and MIT-licensed); Portkey adds compliance and observability tooling; Requesty offers a similar hosted catalog. Together AI and Fireworks AI are a different animal — inference providers hosting open models on their own hardware, not aggregators reselling everyone else’s.

Model families by company

These are per-model API/token prices, not the ecosystem subscriptions above — for live, continuously-updated numbers, see this site’s own LLM Pricing tracker.

Best value by model family, based on LiveBench scores:

• GLM (Zhipu) — value leader among open models: strong scores at a fraction of frontier pricing.
• Qwen (Alibaba) — best for self-hosting: smaller checkpoints run on one consumer GPU, Apache-2.0.
• DeepSeek — best open model overall: trails the closed frontier by a few points, MIT-licensed.

Conclusion

The chatbot and coding-agent space keeps converging on the same middle layers — Markdown, SSE streaming, MCP — while the real differentiation has moved to the edges of the stack: which ecosystem fits how you actually work, and which model gives the best return per dollar. For most builders that settles into a hybrid pattern: a frontier subscription for the work that needs it (Claude for programming and design, ChatGPT for breadth of experience), paired with an open-weight model like Qwen or DeepSeek for the routine, high-volume tasks where self-hosting pays for itself.