MCP Permission Planner
Score shell, file, browser and email MCP exposure before enabling autonomous tools in Odysseus AI.
Free browser tool ยท Updated 2026-11-20
Estimate model weights, KV cache and runtime overhead before pulling large Ollama, llama.cpp or vLLM models into your Odysseus workspace.
Independent guide. Unofficial. Verify hardware drivers and model tags with ollama ps on your own machine.
Instant browser calculator
Adjust parameter count, quantization, context window and parallel agent sessions to avoid silent CPU offloading.
Local inference estimate
Full GPU fit with healthy headroom
Your selected model fits cleanly inside available memory.
Next tools & guides
Move straight from hardware sizing to cost comparison or permission hardening without restarting your search.
Score shell, file, browser and email MCP exposure before enabling autonomous tools in Odysseus AI.
Compare local GPU electricity bills against monthly OpenAI, Anthropic or OpenRouter token spend.
Step-by-step CUDA, ROCm, Apple Silicon and Ollama passthrough checks for Docker and native installs.
FAQ
Odysseus itself uses minimal memory. An 8B Q4_K_M model with 16K context runs well on 8 GB to 12 GB VRAM, 14B Q4_K_M fits 16 GB, and 32B Q4_K_M needs 24 GB or more for full GPU speed.
Longer context windows allocate a larger key-value cache in VRAM. Running 32K to 64K Deep Research prompts or parallel agents can easily double non-weight memory usage.
Ollama and llama.cpp offload overflow layers onto CPU and system RAM. Generation still finishes, but speeds drop significantly compared with 100% GPU placement.
No. You can host Odysseus on any modest computer and point it at OpenAI, Anthropic or OpenRouter, or run smaller quantized models on Apple Silicon unified memory.