Releases·Tool Release
an (unsafe) command shell mcp server
A developer shares cmdshellmcp, a Python shell MCP server that constrains LLM shell access, after IBM Granite 4.2 8B (Q4_K_M) in OpenCode hallucinated tool use and attempted root-path file writes. Running the model through llama-server with the MCP server, the same Flask 'hello world' task completes correctly in one shot.
26 engagement1 sourceTue, Sep 22, 2026, 01:44 AM
AI Overview
- Redditor finds granite 4.2 8B at Q4_K_M quant writes files to root absolute paths and loops hallucinating shell commands instead of executing them in OpenCode
- Running granite 4.2 8B Q4_K_M with the author's cmdshellmcp Python shell MCP server produces the flask hello-world app correctly in one shot
- cmdshellmcp2.py takes an explicit --cwd option to confine commands to a work path, but the author warns the LLM can escape restrictions
- Author recommends running the cmdshellmcp server only inside a throwaway disposable docker container
Related
Products·building
Fine-tuned 1.5B guard model that blocks indirect prompt injection for local tool-using agents
A lightweight fine-tuned Qwen2.5-1.5B-Instruct guard model that detects indirect prompt injection in tool calls for local agents (Ollama, llama.cpp, vLLM with MCP or shell access). It replaces heavier options like Llama Guard 3 or cloud guardrail APIs, which cost VRAM, 150-500ms latency, or privacy leaks. Aimed at developers running local tool-using agents on untrusted content.
A lightweight fine-tuned Qwen2.5-1.5B-Instruct guard model that detects indirect prompt injection in tool calls for local agents (Ollama, llama.cpp, vLLM with MCP or shell access). It replaces heavier options like Llama Guard 3 or cloud guardrail APIs, which cost VRAM, 150-500ms latency, or privacy leaks. Aimed at developers running local tool-using agents on untrusted content.
1 engagement1 sourceredditSat, Aug 29, 2026, 09:07 PM
Social·Community
Redditor runs local Qwen3.6-35B-A3B model in llama.cpp and drives GIMP via maorcc/gimp-mcp MCP server
A Reddit user wired a local uncensored Qwen quantized model into GIMP using llama.cpp's MCP server support and the community gimp-mcp tool, prompting it to draw a flower. The setup works but output quality is poor, with the user attributing results to the small model size.
A Reddit user wired a local uncensored Qwen quantized model into GIMP using llama.cpp's MCP server support and the community gimp-mcp tool, prompting it to draw a flower. The setup works but output quality is poor, with the user attributing results to the small model size.
48 engagement1 sourceredditWed, Sep 16, 2026, 01:52 AM
Social·Community
Users test Meta's Muse Glimmer 30B with Hermes Agent for local coding
A developer tested Meta's Muse Glimmer 30B model locally using llama.cpp and connected it to Hermes Agent, achieving about 22 tokens per second on an M5 Pro. The model performed correct tool calls and produced working code, unlike when run with OpenCode.
A developer tested Meta's Muse Glimmer 30B model locally using llama.cpp and connected it to Hermes Agent, achieving about 22 tokens per second on an M5 Pro. The model performed correct tool calls and produced working code, unlike when run with OpenCode.
10 engagement1 sourceredditTue, Aug 11, 2026, 04:04 PM
Social·Community
Users report Ollama-hosted Qwen3.6 tool calls fail in harnesses; LM Studio works
A user reports that Qwen3.6 models hosted via Ollama fail to execute tool calls in harnesses like oh-my-pi, qwen-code, and opencode, printing tool calls as output instead. The same model works correctly when served via LM Studio, suggesting an Ollama-specific issue.
A user reports that Qwen3.6 models hosted via Ollama fail to execute tool calls in harnesses like oh-my-pi, qwen-code, and opencode, printing tool calls as output instead. The same model works correctly when served via LM Studio, suggesting an Ollama-specific issue.
5 engagement1 sourceredditMon, Jul 27, 2026, 07:02 PM
Social·Community
Users report Muse Glimmer overthinking and tool-call loops in local use
Multiple users report that Muse Glimmer, a local LLM, overcomplicates simple tasks and gets stuck in long terminal tool-call sequences, exhausting context and hitting tool-call limits. The behavior appears specific to Muse Glimmer, as other models like DeepSeek v4 Flash 0731 and Qwen3.6 27B handle similar prompts with far fewer calls.
Multiple users report that Muse Glimmer, a local LLM, overcomplicates simple tasks and gets stuck in long terminal tool-call sequences, exhausting context and hitting tool-call limits. The behavior appears specific to Muse Glimmer, as other models like DeepSeek v4 Flash 0731 and Qwen3.6 27B handle similar prompts with far fewer calls.
010249 engagement3 sourcesredditMon, Aug 10, 2026, 05:51 PM