Claude Code · Codex · Ollama · OpenClaw — one terminal, one transcript.
A multi-peer AI CLI orchestrator with shared memory, cross-checking, parallel swarms, MCP passthrough, and automatic credit fallback.

an AbteeX AI Labs release · site · github · huggingface space

---

DuoX is a drop-in CLI alternative to Claude Code / Codex CLI that runs them together — plus Ollama (local) and OpenClaw (inter-agent channels) — under one supervisor loop. One prompt, many peers, a shared transcript, a single scrolling interface.

Why DuoX?

	Claude Code	Codex CLI	DuoX
Single-agent chat	✓	✓	✓
Multiple agents in one transcript	✗	✗	✓
Cross-validation (planner ↔ executor)	✗	✗	✓
Parallel sub-agent fan-out	✗	✗	✓
Automatic credit/rate-limit fallback	✗	✗	✓
Local Ollama fallback	✗	✗	✓
MCP pass-through	✓	✗	✓
CLAUDE.md / AGENTS.md auto-context	✓	✓	✓
@file mentions, /compact, /remember	✓	partial	✓
Headless -p + JSON output	✓	✓	✓
HTTP API daemon	✗	✗	✓

Install

pip install duox

From source:

git clone https://github.com/aimaghsoodi/duo && cd duo
pip install -e .

Requires Python 3.10+. After install:

duox init      # write ~/.duo/config.toml + sample skill
duox doctor    # verify peers, MCP, readline, openclaw

duox doctor tells you exactly what's missing and how to fix it.

Quick start

Interactive (default — Claude-Code-like loop):

duox
# then type your goal, or any of:
#   /help  /tips  /examples  /peers  /cost  /compact  /remember  /doctor
#   @src/main.py                  (inline file excerpt)
#   @src/main.py:40-80            (line range)
#   /cd ../other-repo             (switch working dir mid-session)

Headless (for scripts / CI / git hooks):

duox -p "summarise the last 10 commits"
duox -p "run the test suite and fix any failures" --output-format json
echo "review my changes" | duox -p

With specific peers / supervisor:

duox --peers claude,codex,ollama --supervisor claude "refactor the auth module"
duox --parallel "run these three benchmarks"
duox --resume abc123   # resume a previous session

Try it live

A Gradio demo of the orchestration loop is hosted on Hugging Face:

👉 https://huggingface.co/spaces/AbteeXAILabs/duox

Features

• Shared transcript — every peer sees every message.
• Supervisor / executor loop — one peer plans & validates, others execute.
• Parallel swarm mode — fan the same task across peers, supervisor merges.
• Credit fallback — rate-limit/429/"out of credits" auto-fails over.
• Context files — CLAUDE.md, AGENTS.md, DUO.md auto-loaded (walks up from cwd).
• @file mentions — inline file/dir/line-range excerpts in any prompt.
• Skills — markdown prompts in ~/.duo/skills/ auto-injected by keyword match.
• Sessions — JSONL event log + transcript under ~/.duo/sessions/; duox --resume.
• Hooks — run commands on pre_step/post_step (tests, linters, formatters).
• MCP pass-through — [mcp.servers.*] in your config is forwarded to Claude.
• HTTP API — duox serve --port 8787 exposes goals/sessions over JSON.
• Per-project config — drop a .duo.toml in any repo to override globals.
• Line editor — history, ↑/↓, tab-complete on slash commands + peer names.
• 30+ slash commands — /help lists them all.

Configuration

~/.duo/config.toml:

peers        = ["claude", "codex", "ollama"]
supervisor   = "claude"
ollama_model = "llama3.1"
max_steps    = 30
parallel_default = false

[hooks]
pre_step  = []
post_step = ["pytest -q"]

[mcp.servers.context7]
command = "npx"
args    = ["-y", "@upstash/context7-mcp"]

[openclaw]
thinking = "high"

Project-level .duo.toml in your repo root is deep-merged on top.

Headless / scripting

# Plain text (stdout = final answer)
duox -p "give me a release-notes draft"

# JSON envelope with per-peer stats
duox -p "audit this PR" --output-format json | jq '.final'

JSON output shape:

{
  "ok": true,
  "final": "…final answer…",
  "session_id": "20260418-abc",
  "turns": 14,
  "peers": {"claude": {"calls": 4, "seconds": 12.3, "tokens_in": 4280, "tokens_out": 910}}
}

Architecture

user prompt
    ↓
 supervisor ──► JSON decision {action, to, instructions}
    │
    ├── delegate ─► executor peer(s) ─► output
    │                     │
    │                     └── (parallel) ─► merge
    ├── execute_self
    ├── validate
    └── done  ─► final answer

Under the hood: claude -p --dangerously-skip-permissions, codex exec --dangerously-bypass-approvals-and-sandbox, ollama run, and openclaw agent, streamed through a small Python orchestrator.

Note: the internal Python package is still named duo for backwards compatibility — from duo import … continues to work. The user-facing CLI is duox, with duo kept as a legacy alias.

Safety

DuoX runs peer CLIs with permissive flags so they can actually do work. Run it in a directory you trust — ideally a scratch repo, container, or VM. The transcript shows every action; read it.

Develop

pip install -e '.[test]'
pytest -q          # 47 tests, 3 OS × 3 Python versions in CI

Links

• PyPI: https://pypi.org/project/duox/
• GitHub: https://github.com/aimaghsoodi/duo
• HF Space (demo): https://huggingface.co/spaces/AbteeXAILabs/duox
• HF model card: https://huggingface.co/AbteeXAILabs/duox
• Issues: https://github.com/aimaghsoodi/duo/issues

License

MIT © 2026 AbteeX AI Labs.