“I wouldn't run a heavy harness like claude code with that amount of vram. The system prompt eats all the memory before you even start giving it actual directions and code to look at. I have been using pi, but there are a lot of light weight harnesses that won't use all your vram to cache the system prompt.”
“Yoetz is my host-owned receipt layer for Claude Code (also Cursor / Codex): tool calls / checks / file touches as claim-vs-did, so “done” isn’t transcript vibes. complementary to what you built, not a replacement for the runner.”
“I've been experimenting with different models, some local, some through OpenRouter, and I had Claude Code as the driver, whatever orchestrator, whatever you want to call it. It was quick. It snapped back pretty quick on Telegram.”
“One snag I hit: I don't see a way to disable the local sandbox option in Claude Code settings. It says that the setting is 'managed by my organization.' These are my personal Max plans, and there is no organization.”
“This isn't a popular answer, but instead of complaining about how bad Claude codes, write the code yourself so it's commented the way you want and uses code you want to use. Then, allow Claude to review, check, recommend, etc.”
“One thing that I didn't understand from this article was how you use this deployment platform.Can you tell Claude code for web to build and deploy a web app using it? If so, does that end up with a public URL somewhere? Does that deployment continue working after your Claude code for web session ends?”
“the biggest trap when moving away from claude code is overengineering the local harness on day one. starting minimal with something like pi or taking inspiration from its lean setup gives you clear visibility over context and token usage, especially when paired directly with llama.cpp inside tmux instead of stacking heavy frameworks.”
“The OP is on a vRAM-constrained system, opencode,claude code just bloat the hell of the context just with the system prompt. I did not recommend barebones [pi.dev](http://pi.dev), but I added the plugins I have been using for weeks without any issues.”
“Claude Code relies on massive context windows and near-zero tool schema degradation. With 24GB VRAM, your sweet spot model is \*\*Qwen 2.5 Coder 32B Instruct (quantized to Q4\_K\_M or IQ4\_XS)\*\* via \`llama.cpp\` / \`vLLM\` / \`Ollama\`.”
“That said, if someone's use case is "I'd like to know, what harness, right out of the box would feel most at home for current Claude Code users" I'd probably point them to opencode.”
“Migrating off Claude Code usually means reinventing the harness: shell access, file state, retries, and a place the agent can fail and recover. Local is great for privacy, but the ops tax adds up fast.”
“Aider and Continue both feel close to "chat over the repo" if you already vibe in Claude Code; OpenCode / similar TUI tools are closer if you liked the terminal loop.”
“Over the last year I've gotten quite comfortable with claude code, and it seems likely there were be a gradual cost rug pull, and I'd like to put myself in a better position when that happens for local use. I am already used to running local models (such as Qwen3.8_Q5) in things like lmstudio, but I have no experience with harnesses.”
“The problem here is the statement ‘Claude Code is the best developer I e ever worked with.’ If that’s true you haven’t worked with the right developers. There’s another comment here about AI augmenting people with expertise.”
“I've been using Claude Code for less than a month, and I'm blown away. I've rebuilt a database app in less than a month that took 3 years and over 12k to build. It's the best coder I've ever worked with.”
“Then it directs your request to the next available provider, let's say GPT, so the request, with the context, with the prompt and everything else now reaches OpenAI model, which will reply with a set of instructions/tools (for example, use edit tool on file X to change/add/remove Y lines) + an answer + many other things that get translated and mapped in my software to be same as it was coming from Anthropic and then reach the claude code agent "engine" that knows how to handle all of it.”
“Do you use coding agents? Just curious bc from my experience using coding agents, open ai’s codex is neck and neck with anthropic’s claude code if not ahead. I wouldn’t agree that OpenAI hasn’t done anything since ChatGPT since codex is my daily driver for software engineering”
“Much saner English, much better at execution, and gets stuff done in a matter-of-factly kind of way (Claude Code is a mess these days -- it gets things wrong and goes around in circles).But I'm harness agnostic and am not locked in.”
“Vibe is noticeably cheaper per task but it loses the thread on anything over about four files, which is where Claude Code still earns its tokens. I ran both on the same refactor in October: Vibe was roughly a third of the cost and took two extra rounds to get it right, so call it a wash.”
“I run four tmux sessions with 20-30 Claude Code chats each, across 30-plus Go services, a mobile app, infra repos and a data pipeline. The agents cope fine. I didn't.”
“Like using opusplan in Claude code where the planning is done with a more capable model and the work with a more cheap model? Can you please explain how is your workflow?”
“How will it keep the data if let’s say I’m on Claude code session, and then switch to GPT after usage runs out? You mention it’s on the same terminal, but how will codex take over?”
“I built most of it with Claude Code. The pattern every night was the same. One session holds the plan and only delegates, the sessions doing the work each get their own git worktree so no two ever touch the same file, hooks block a message from the wrong sender before the model reads it, and CI is the only gate that counts.”
“Need of the hour, with Claude code subscriptions being allocated by the organisation based on project priority. this bring your own subscription is thoughtful.”
“Claude Code has started giving me private asides when it has background agent workflows running, such as: Privately: the memory guard flagged my new file as an orphan, so it needs an index line in a topic file, and I need to see that file's sections to place it correctly.”
“Advice written for the claude.ai zip upload won’t fix a Claude Code “skill not found,” and vice versa. Say which surface you’re on before troubleshooting.”
“We are team of 8 people. Everyone is fresh grad, inexperienced with not an ounce of real world engineering knowledge (best practices, proper workflow). My entire company and I are new to agentic coding and to be honest, we are kind of struggling with development.”
“Claude Code did most of the heavy lifting on the server, the three MCP tools, and the tests. The pairing and expiry logic got rewritten a few times with Claude's help before it finally felt right.”
“I'm not sure how is it with personal subscription, but being forced to use Claude Code at work, I've never hit any limit, including long sessions with Opus”
“I've never used Claude Code, only heard that you hit the limit really fast, if you actually use the fancy models you burn through allotment fsr too quickly.”