DeskPilot: open-source native Python desktop client for local LLMs with MCP and sandboxed tools
poofph · reddit · 2026-09-23
A developer open-sourced DeskPilot, a native Python/Tkinter desktop AI client (built with heavy help from Qwen 3.8 27B) with a security-first agentic architecture for local LLMs.
Highlights:
- Works with Ollama, LM Studio, vLLM, llama.cpp, or any OpenAI-compatible endpoint
- Built-in dependency-free stdio JSON-RPC 2.0 MCP client; dynamically converts MCP server tools into OpenAI function-calling schemas
- Per-tool permission bar (Always Allow / Ask / Off), unvetted tools default to ask
- Built-in tools: SearXNG/Exa search, Firecrawl scraping, file I/O with PDF, clipboard, screenshots, sandboxed Node.js execution
- Streaming reasoning display with timing ("Thought for X.Xs")
- 100% local voice: faster-whisper STT and kokoro-onnx TTS with a look-ahead pipeline
- Context metrics (decode tok/s, TTFT, usage gauge) and automatic structured handoff summaries at a configurable context threshold
- Security guardrails: SSRF/DNS-rebinding protection, filesystem sandboxing, daemon threading for a fluid UI
Runs on Python 3.10+ or compiles to a single .exe via PyInstaller. GitHub: shuhdonk/DeskPilot.
More from coding & agent
- AI engineering is more like lawmaking than board games, argues Drew Breunig — dbreunig · 2026-09-23
- Seroter's daily digest: GPT-6 and Opus 5.5 ship, 1 in 4 agents run unmonitored — rseroter · 2026-09-23
- Spawning Claude agents that auto-open terminal panes: 'tmux can't do this' — letandrewcook · 2026-09-23
- Toddler's interactive storybook built in two hours with an agent workflow — mimi10v3 · 2026-09-23
- Your data stack is about to get less forgiving: agents need data that's true now — bigdata · 2026-09-23
- Opinion: Agents make software good at using software, not just being used — r0ck3t23 · 2026-09-23