DevMind
An AI software engineering agent that reads, modifies, tests and builds real code.
01overview
DevMind is an experiment in automated software engineering: an agent that takes a task, explores a codebase, and reasons about the changes required to implement, fix, or extend it.
It is built around an LLM core (Qwen through Ollama, routed via LiteLLM) exposed through a FastAPI + WebSocket interface with streaming responses, so the agent's reasoning and tool calls can be observed as they happen.
A Flutter client provides the operator interface, while credentials and secrets are kept out of plain configuration through encrypted storage. The project is an evolving platform for LLM agent workflows, not a finished product.
02architecture
- 01▣
Task Input
natural language / issue
- 02▣
Agent Loop
plan · reason · act
- 03▣
Tool Calling
read · search · edit · run
- 04▣
Codebase Operations
modify · test · verify
- 05▣
Streaming Output
WebSocket · FastAPI
03stack
05software
- Python
- FastAPI
- WebSocket
- Ollama
- Qwen
- LiteLLM
- AI agent tool calling
- Streaming protocols
- Encrypted credential storage
06problems & solutions
Grounding an LLM agent in the actual state of a codebase rather than hallucinated paths.
Tool-based file search and read operations feed the model real context before it proposes edits. Iteration is being documented as the agent matures.
Streaming long agent transcripts without blocking the interface.
FastAPI + WebSocket streaming keeps reasoning, tool calls and diffs flowing incrementally to the Flutter client.
07results & specs
- Interface
- Flutter client + streaming transcript
- Model runtime
- Ollama / Qwen via LiteLLM
- Metrics
- Documentation in progress.
08gallery & media
01