// coding agent

A coding agent that runs on the GPU you already own.

Hubbleflow reads and edits files, runs commands, searches the web and keeps a todo list, in your terminal, against whichever model you point it at. No API key and no account to start.

The Hubbleflow terminal listing local models, Mesh and Ollama, with Gemma4:12b selected

Install it

macOS and Linux:

./install.sh

Windows, in PowerShell:

.\install.ps1

The script installs uv if you don’t have it, puts hubbleflow on your PATH, and creates ~/.hubbleflow/.env for your keys. uv fetches its own Python, so you don’t need 3.13 installed first. Add --dev to install editable if you plan to work on the harness itself.

Run it

cd your-project
hubbleflow

No key is required to start. A local Ollama model or a mesh node runs without one. For Gemini put GOOGLE_API_KEY in ~/.hubbleflow/.env; for NVIDIA, NVIDIA_API_KEY.

Point it at a model

Seven providers, discovered live rather than from a hardcoded list. Type a bare model name and the harness works out where it lives.

/models          everything reachable right now
/local           Ollama, vLLM, llama.cpp, FreeToken, Mesh
/model gemma4:12b

Local servers need nothing but a running process. Start ollama serve, vllm serve or llama-server, and it appears on the next listing.

What it does

Reads and edits files inside your workspace, and runs commands - reads are free, writes and commands pause and ask.
Picks up AGENTS.md or CLAUDE.md from the project root, so its conventions don’t need re-explaining each session.
Connects to MCP servers declared in .mcp.json, in the same shape that Claude Code uses.
Researches a question over several rounds and files what it finds into a knowledge bundle, so the next session already knows it.

Remove it

./uninstall.sh            # removes the command, keeps your config
./uninstall.sh --purge    # also deletes ~/.hubbleflow

Apache 2.0, and the source is on GitHub.