Start fast, resume anytimesession-aware
Launch interactive, run one-shot, or pick up exactly where you left off. Sessions are scoped per project directory.
built by aktagon.com
Quantum is a minimal coding agent for your terminal — a Claude Code-style console with slash commands, per-project sessions, and an approval gate on every tool call. Generated into Python, Rust, TypeScript, and Go — not hand-ported — with the core verified identical. Providers handled by llmkit.
Eleven providers. Four languages. One loop you own.
$ quantum quantum — coding agent · provider: anthropic · model: claude-opus-4-8 ~/repo › add a /healthz endpoint to the server ● edit server.py (+11 −0) approve? [y/N] y ✓ applied · 1 file changed ~/repo › /cost steps: 3 · tokens: 18.4k · cost: $0.14 ~/repo › /provider ollama qwen2.5-coder switched → ollama · qwen2.5-coder
Bring your own model. · Generated, not locked. · Sandboxed by the kernel.
What it does
Launch in a directory, talk to the agent, approve its edits. Switch models or providers mid-session. Sessions persist per project.
Launch interactive, run one-shot, or pick up exactly where you left off. Sessions are scoped per project directory.
Quantum is generated into Python, Rust, TypeScript, and Go — not hand-ported per language — and the deterministic core is verified identical across all four by a shared conformance suite that runs in CI, so there's no drift. Python is the reference runtime today; the other cores are gated now.
Everything runs inside the interactive console — switch models, inspect cost, manage context, resume work, without leaving the prompt.
One agent, eleven backends — frontier APIs or a local server on loopback. The provider plumbing is handled by llmkit.
The Task tool spawns a subagent that runs in its own isolated context and returns only its final message — a frontier model orchestrates while a focused subagent does the work, optionally on a different or cheaper model. Subagents can be named markdown agents; quantum reads the same ~/.claude/agents format.