Andromity

A terminal AI coding agent. Autonomous by choice, gated by trust.

English | 简体中文 | Русский | Português (Brasil) | 日本語 | Deutsch | Français | Español | हिन्दी

Install

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/agenticmarket/andromity/main/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/agenticmarket/andromity/main/install.ps1 | iex

# Or with pipx
pipx install andromity

Needs Python 3.11+. The installer handles pipx if you don't have it.

Then:

andromity

Point it at a folder, answer the trust prompt, pick a model — you're in.

Andromity is not a chat window. Not a plugin. It's a proper terminal workspace — sessions, diffs, file viewer, cron scheduler, profiles — with an AI agent wired into all of it.

It works with Claude, Gemini, GPT, Groq, Ollama, and anything LiteLLM supports. Run it entirely local with Ollama and pay nothing. Swap models mid-session with Ctrl+L.

The thing that makes it different: nothing runs until you say the folder is trusted.

How the trust model works

When you open a folder, Andromity asks if you trust it. That answer controls everything — not your permission mode, not your API key, not your settings. Say no, and the agent cannot write a file, run a command, or touch anything. Full stop.

Say yes, and you pick how much rope the agent gets:

Start in SAFE. Move to YOLO when you know what the agent does in your codebase. /trust and /untrust any time.

Headless / scripted

andromity run "add error handling to auth.py"
andromity run "refactor this to async" --yes      # auto-approve everything
andromity run "review session.py" --dry-run       # see what it would do

What's inside

Scheduler. Run the agent on a timer while you sleep. /cron opens the scheduler. Jobs persist per project in .andromity/crons.json. Works with any permission mode — use YOLO for fully unattended runs.

Profiles. Switch what the agent is trying to do.

• builder — plans, then implements
• coder — implements directly, no planning phase
• reviewer — read-only, produces findings
• planner — plans only, touches nothing

MCP support. Drop a mcp.json in your project. Tool schemas index first, full payloads load only when the agent needs them. Keeps token use sane with 50+ tools connected.

Sessions. Everything is saved. Switch between sessions with /sessions or Ctrl+O. /compact when context gets heavy. /undo to revert the last turn and all its file changes.

Sound notifications. The agent pings you when it needs approval or finishes a turn. Toggle them independently under Ctrl+E → Advanced → Sounds.

Model-agnostic. LiteLLM under the hood. Anthropic, OpenAI, Gemini, Groq, OpenRouter, Ollama, NVIDIA NIM. Switch mid-session with Ctrl+L.

How it compares

Privacy

Your code goes to one place: the LLM provider you configure. Not us.

• API keys live in ~/.andromity/config.toml
• Sessions stored locally in ~/.andromity/sessions/
• Anonymous ping on first launch — no code, no paths, no keys. Full details in telemetry-worker/README.md
• Opt out: export DO_NOT_TRACK=1, or telemetry = false in config, or Ctrl+E → Advanced → Telemetry

💡 Run /help inside the workspace. Not every feature is documented here — discovery is part of the experience.

Contributing

Open an issue or PR. Honest feedback and bug reports are more useful than feature requests right now.

See CONTRIBUTING.md for project layout and dev setup.

MIT — see LICENSE.