› /plan add retry with exponential backoff to fetch_json() in @orbit/net.py
• project_overview(path='.')
• read_file(path='orbit/net.py')
✓ proposed plan, approved
• multi_edit(path='orbit/net.py', edits=[2 items])
+ for attempt in range(retries + 1):
• run_command(command='pytest -q')
✓ 2 passed
Geocentric Code
A coding agent that stays on your machine.
Geocentric Code plans, reads, edits, runs and verifies from your terminal. The model runs locally through Ollama. You need no API key, you pay no per-token bill, and nothing you write leaves your computer.
- MIT licensed
- macOS, Linux, Windows
- 22 tools
- Python 3.9+
Install
One command (then you are done)
The installer adds git, Python, Ollama and ripgrep if you lack them, downloads the code, and fetches a model sized to your memory. It asks before it installs anything.
macOS & Linux
curl -fsSL https://geocentricai.com/install.sh | bashOpen the script first if you like. Run it again any time to update.
Windows (PowerShell)
irm https://geocentricai.com/install.ps1 | iexOpen the script first if you like. Run it again any time to update.
01Checks your machine
Looks for git, Python 3.9 or newer, Ollama and ripgrep. Anything missing is installed with Homebrew, your Linux package manager, or winget.
02Fetches the code
Clones geocentric-code from GitHub into a private folder and pulls only the parts it needs.
03Builds a private environment
Creates a virtualenv and installs the dependencies into it, so nothing touches your system Python.
04Puts geo on your PATH
Adds the command to ~/.local/bin on macOS and Linux, or to a per-user folder on Windows.
05Downloads a model
Picks one for your memory: 14B at 16 GB or more, 7B at 8 GB or more, 3B below that. It skips this step if Ollama already has models.
Installer options
Change what it does
| Option | macOS & Linux | Windows |
|---|---|---|
| Skip the model download | --no-model | $env:GEO_NO_MODEL='1' |
| Also install the web gateway | --gateway | $env:GEO_GATEWAY='1' |
| Never ask for confirmation | --yes | $env:GEO_YES='1' |
| Show the plan, change nothing | --dry-run | $env:GEO_DRY_RUN='1' |
| Remove it (keeps Ollama and models) | --uninstall | $env:GEO_UNINSTALL='1' |
On macOS and Linux, pass options after bash -s --, for example curl -fsSL https://geocentricai.com/install.sh | bash -s -- --no-model. On Windows, set the variable in the same window, then run the install command.
Plan, edit, verify.
Local by default
Models run through Ollama on your own hardware. Once a model is pulled, nothing needs the network.
Plan mode
/plan makes the agent survey your codebase in read-only mode and return a structured plan. You accept it, ask for changes, or cancel.
22 tools
Read, write, exact-string and multi-hunk edits, search, shell, Python, background jobs and a full git set. Independent reads run in parallel.
Guardrails you can see
Writes stay inside your working directory. Credential-like files always prompt. Destructive commands ask even with auto-accept on.
Works with the model you have
Models without native tool calling get tools described in the prompt. Fake tool results are flagged so you do not trust invented output.
Sessions and custom commands
Every session autosaves, resumes and exports to markdown. A markdown file in .geocentric/commands/ becomes a slash command.
Start in any folder.
geointeractive session in the current directorygeo "add a healthcheck endpoint"run one prompt and exitgeo -cresume the most recent session heregeo --planstart in read-only planning modegeo -m qwen3-coder:30bpin a model for this runRead the source.
Geocentric Code is MIT licensed and developed in the open. Open an issue, read the code, or send a pull request.