andromity
ANDROMITY · v0.2.3

A terminal AI
coding agent.
Gated by trust.

Not a chat window. Not a plugin. A proper terminal workspace — sessions, diffs, file viewer, cron scheduler, profiles — with an AI agent wired into all of it.

Works with Claude, Gemini, GPT, Groq, Ollama, and anything LiteLLM supports. Swap models mid-session with Ctrl+L.

WINDOWS (POWERSHELL)
$ irm https://raw.githubusercontent.com/agenticmarket/andromity/main/install.ps1 | iex
Needs Python 3.11+. Point it at a folder, answer the trust prompt, pick a model — you're in.
02 · TRUST MODEL

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.

SAFE Approve each one
  • Plans: Approve each one
  • File writes: Approve each one
  • Shell commands: Approve each one
TRUST Review plans only
  • Plans: Approve
  • File writes: Direct (no review)
  • Shell commands: Direct (no review)
FULL Direct execution
  • Plans: Auto
  • File writes: Direct
  • Shell commands: Direct
YOLO Fully unattended
  • Plans: Auto (FYI only)
  • File writes: Silent
  • Shell commands: Silent

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

03 · CORE FEATURES

What's inside

Everything needed for real coding sessions, wired directly into a proper terminal workspace.

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.

Agent 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 & /undo

Everything is saved. Switch between sessions with /sessions or Ctrl+O. Use /compact when context gets heavy and /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.

04 · CLI AUTOMATION

Headless / scripted

Automate refactors, error handling, and code reviews straight from your shell or CI/CD pipelines.

Direct task execution 
andromity run "add error handling to auth.py"
Auto-approve everything (unattended mode) 
andromity run "refactor this to async" --yes
Dry run inspection (preview without modifying) 
andromity run "review session.py" --dry-run
05 · BENCHMARK

How it compares

Built from the ground up around folder trust, cron scheduling, and terminal ergonomics.

Feature Andromity Aider Cursor Claude Code
Folder trust model
Permission levels (SAFE → YOLO) Partial
Built-in cron scheduler
Inline diff viewer
Session management + /undo Partial
Agent profiles Partial
Local-first / Ollama / BYOK
MCP support Partial
TUI workspace
06 · PRIVACY FIRST

Privacy

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

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