quantum

A coding agent that
asks before it acts.

quantum is a minimal, multi-provider coding agent for your terminal. Every command is approved before it runs, every tool call is sandboxed at the kernel on macOS, and every session survives a restart. Python, MIT-licensed.

> install quantum ⏎ View source
quantum — zsh
$ quantum
quantum — anthropic/claude-sonnet-4-6  (/help for commands)
> fix the failing test in tests/test_config.py

# Run the failing test to see the error
$ pytest tests/test_config.py -q
Approve? [y] yes  [a] all  [n] no: y

The test expected load_config to fall back to the provider default
when the project config omits a model. Fixed the fallback and the
test passes.
[session: 4 steps · 18,204+1,310 tok (12,876 cached)]
>

  • Asks before it runs

    Shell commands are shown in full — command, cwd, timeout — and wait for a yes, an all, or a no. Read-only tools skip the prompt and run on a read-only profile instead.

  • Sandboxed at the kernel

    On macOS every tool call runs under a SeatBelt profile: writes are contained to the directory you launched from, secret paths like ~/.ssh are unreadable, and read-only tools get no network at all.

  • Sessions that survive

    Conversations are stored per project. quantum -c continues the most recent one, -r picks from a list, and /resume switches mid-conversation.

  • Any provider

    Anthropic, OpenAI, Google, Groq, Grok, OpenRouter — or a local Ollama, vLLM, or llama.cpp server with no API key. Switch with /provider without losing the session.

  • A real console

    Line editing and history, live model listings with /model, token and cost accounting with /cost, context compaction with /compact, and hard budget caps on turns, time, and spend.

Install

Requires Python 3.10+ and uv.

$ git clone https://github.com/aktagon/quantum
$ cd quantum && make install
$ export ANTHROPIC_API_KEY=…
$ quantum

Local providers with a loopback address — Ollama, vLLM, llama.cpp, LM Studio, Jan — need no key.