Puzld.Ai

The interesting thing about PuzldAI is what it refuses to do: it does not host models, sell tokens, or ask you for a stack of API keys. It sits on top of the CLI tools already on your machine — Claude, Gemini, Codex, Ollama, Mistral — and supplies the orchestration layer those tools individually lack. That architectural choice defines both its appeal and its ceiling. The appeal is real. Intelligent routing sends prompts to a model based on task complexity and type, so a trivial rename does not burn your most expensive agent. Compare mode puts responses side by side; Debate mode has models critique one another; Correction mode iterates on a weak output; Consensus mode drives toward agreement. For teams arguing about which model writes better migrations, that is a genuinely useful laboratory, and the vendor's own framing — "like Claude Code, but for any LLM" — is a fair description of Plan and Build modes, which give models file editing, bash commands and code generation against your actual repository. The secondary toolchain is where the engineering depth shows. Persistent memory with vector search injects relevant context automatically rather than making you re-paste files. AST indexing, dependency graphs and semantic code search mean the agent can navigate a large codebase structurally instead of grepping blindly. Pipelines chain agents into multi-step workflows, /autopilot runs autonomously, and the same machinery can be pointed at training-data generation — a niche but demanding use case that needs exactly this kind of scripted multi-model collaboration. The ceilings are worth naming. You need terminal fluency; the tool exposes TUI commands and CLI commands like puzldai run, puzldai compare and puzldai agent, and that is the whole interface. It is not cloud-hosted, so nothing runs when your machine is off, and every capability depends on the wrapped CLI tool being present and working. The router defaults to a fast local model via Ollama but can be pointed at Claude or Gemini, and if the chosen agent fails it falls back to a configured default (Claude unless changed) — useful resilience, but it also means your effective model mix is a configuration decision you own. Where it fits: developers who already run several coding CLIs, teams that want evidence before standardising on one model, and researchers building multi-agent training data. Where it does not: anyone who wants a GUI, a hosted runtime, or a single opinionated agent with a polished onboarding path. If you only ever use one model, the orchestration overhead buys you little — Claude Code or Aider will feel more direct. If you use three, PuzldAI is the only thing in this comparison that treats them as a team.