Skip to content

About

Desktop AI assistant for programmers on local LLMs: Qt / C++, command protocol, own C++ RAG, tree-sitter project graph

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

DevAssistant

Build

English · Русский

A desktop AI assistant for programmers that runs on local models and works on your project through a fixed set of commands, so you always see what it is about to change. Qt 5.15 / C++17, Windows.

Main window

What it does

  • Works on the project through commands, not free-form edits. The model answers with JSON commands — open a file region, edit a line range, apply a unified diff, search the project, run a shell command, ask the user — and DevAssistant executes them. 36 commands, see docs/COMMANDS.md.
  • Safe by default. Every path must stay inside the project folder; each step of a workflow lists the commands it may use; read-only commands run at once, and anything that changes files, runs a program or deletes is shown first with the exact command line, paths and line ranges. Changes are applied all-or-nothing and can be undone.
  • Local models. Ollama or any OpenAI-compatible server (llama.cpp, LM Studio, vLLM).
  • Questions about your documents (RAG), written in C++. Index a folder (code, notes, Markdown, PDF): chunks and their embeddings are stored in SQLite next to the project and updated incrementally; answers cite the excerpts they use. The agent can search the index itself with SEARCH_DOCS.
  • LightRAG client. Upload documents to a LightRAG server, ask in its graph or vector modes and explore the knowledge graph.
  • Code insight. Built-in editor with highlighting, the syntax tree from tree-sitter, and a project graph of classes, methods and calls across files (C/C++, Python, Dart).
  • Workflows. Multi-step prompt sequences with per-step commands, grouped into solution templates; an optional agent mode plans each step with a separate model call.
  • Voice. Dictate to the chat with faster-whisper, Vosk, whisper.cpp or a wav2vec2 model; have answers read aloud.
  • English and Russian interface.
Confirmation of the model's commands Questions about documents
Command confirmation Documents

Quick start

  1. Download DevAssistant-…-windows-x64.zip from Releases, unpack it and run DevAssistantQt.exe. Nothing to install.
  2. Install Ollama and get a model, for example ollama pull qwen2.5-coder:7b (and ollama pull nomic-embed-text for document search). Another OpenAI-compatible server can be chosen in Settings › Model.
  3. Open a project folder, pick the model in the chat and describe the task.

Voice input and text to speech are optional: they need Python 3.10+ and pip install -r scripts/requirements.txt, then Settings › Audio.

How a step works

flowchart LR
    Q[Request] --> P[Prompt with the step's<br/>allowed commands]
    P --> M[Local model]
    M --> J[JSON commands]
    J --> G{Policy:<br/>inside the project?<br/>allowed in this step?}
    G -->|no| B[Reason back to the model]
    G -->|yes| C[Confirmation<br/>for changes]
    C --> A[Apply as one batch<br/>with undo]
    A --> R[Results back<br/>to the model]
Loading

More in docs/ARCHITECTURE.md.

Building from source

Requirements: Qt 5.15 with MinGW 8.1 64-bit, Git, PowerShell.

git clone https://github.com/TonyLex98/DevAssistant.git
cd DevAssistant
powershell -ExecutionPolicy Bypass -File scripts/fetch_tree_sitter.ps1
mkdir build; cd build
qmake ../DevAssistant.pro CONFIG+=release
mingw32-make -j4
mingw32-make check

This builds the core library, the program (build/bin/DevAssistantQt.exe) and the tests, then runs the tests. fetch_tree_sitter.ps1 downloads tree-sitter and the C++, Python and Dart grammars at pinned commits. To build the Windows package: scripts/deploy_windows.ps1 -ExePath build/bin/DevAssistantQt.exe -Zip DevAssistant.zip.

Tests (QtTest, 12 suites): the agent engine end to end with a scripted model on a temporary project, the model client against a fake HTTP server, the session, path sandbox, command parser and policy, file commands, change batches with undo, line-ending-preserving edits, prompt data, the code indexer.

CI (GitHub Actions) builds, runs the tests, updates the Russian translation, packages the zip, starts the program headless to take the screenshots above, and publishes releases.

Project layout

Folder Contents
src/core everything the assistant does, as a static library without QtWidgets: session (the session state), model (asynchronous model client), workflow (the agent engine, sequences, templates), commands, context, project, code (tree-sitter index), rag, util
src/ui the main window, chat, dialogs, editor, graphs, document windows
src/app main() and resources
src/audio speech recognition engines, text to speech
third_party md4c; tree-sitter is downloaded here
tests QtTest suites, linked against the core library
translations Russian translation

The layers and the flow of one step are described in docs/ARCHITECTURE.md.

License

MIT, see LICENSE. Third-party components: THIRD_PARTY_NOTICES.md.

About

Desktop AI assistant for programmers on local LLMs: Qt / C++, command protocol, own C++ RAG, tree-sitter project graph

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages