Powerful Features, Zero Cloud Dependency
Interactive Chat
Natural language conversations with full context awareness of your codebase.
Smart Code Completion
Context-aware suggestions powered by your local codebase.
Natural Language to Shell
Describe what you need and get the perfect shell command.
Error Explanation
Understand and fix errors faster with AI-powered explanations.
Code Refactoring
Safe, AST-aware transformations across your entire codebase.
100% Offline
Your code never leaves your machine unless you explicitly opt-in.
Core Features (v1)
| Command | Purpose | Notes |
|---|---|---|
| velo chat | Multi-turn REPL / agent | MCP 0.1 + OpenAI function-calling; auto-detects Qwen3, Osmosis-MCP-4B, smolLM2. |
| velo complete | Code completion at FILE:LINE | Hybrid retrieval → DSPy pipeline. |
| velo shell | NL → shell (-y auto-run) | Works with Bash, zsh, MSYS, PowerShell. |
| velo explain | Explain error logs | Summarise + suggest fixes. |
| velo refactor | Repo-wide refactor | AST-aware patching; optional quality rerank. |
Architecture
LLM runtime
llama-server Vulkan GPU off-loading (-DGGML_USE_VULKAN=ON / LLAMA_VULKAN=1). Default API http://127.0.0.1:8080/v1/…; built-in Web UI http://127.0.0.1:8080/; automatic CPU fallback; optional cloud via --cloud / VELO_CLOUD=1.
Orchestration
DSPy modules. Declarative pipelines; BootlegOptimizer fine-tuning.
Agent layer
MCP 0.1 + Qwen-Agent templates. Structured tool calls, code-interpreter, browser automation.
Retrieval
Two-stage RAG (HNSW file → Tree-Sitter chunks). < 50 ms retrieval; cross-encoder rerank optional.
Data & Storage
Embeddings
E5-large-v2 (384-d) per repo.
Cache layout
~/.velo/ (models, coarse index) + .velo/ (fine chunks, checkpoints).
