← sources

PLEASE HELP Why do local AI coding agents struggle so much with using tools?

source: reddit (reddit:ollama)published 2026-10-31 20:24ingested 2026-11-01 02:40

retained text

I’ve been experimenting with local AI coding agents for a while now, mainly using Ollama with different Qwen models.  
I’ve tried OpenCode, Claude Code, and now Codex CLI. I’ve also experimented with connecting AI to Roblox Studio through MCP.  
The weird thing is that the models can have perfectly normal conversations and write code, but as soon as I ask them to actually do something, like read a file, run a command, or use an MCP tool, things start falling apart.  
Sometimes they just tell me they don’t have access to tools. Other times they literally print out tool-call syntax instead of executing it.  
I’ve tried different models, permissions, configurations, and agents, but I keep running into similar problems.  
Is this just how local models are right now, or am I doing something wrong? Has anyone managed to get a genuinely reliable local AI coding agent working?

document fetched 2026-11-01 00:37 · canonical https://reddit.com/r/ollama/comments/1xkkg30/please_help_why_do_local_ai_coding_agents