Skip to main content

DuoX — multi-peer AI CLI orchestrator (Claude + Codex + Ollama + OpenClaw). Shared transcript, parallel swarms, cross-checking, credit fallback, MCP, headless mode. By AbteeX AI Labs.

Project description

DuoX — by AbteeX AI Labs

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

PyPI Python CI License


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 filesCLAUDE.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 APIduox 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

License

MIT © 2026 AbteeX AI Labs.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

duox-0.2.6.tar.gz (44.1 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

duox-0.2.6-py3-none-any.whl (42.9 kB view details)

Uploaded Python 3

File details

Details for the file duox-0.2.6.tar.gz.

File metadata

  • Download URL: duox-0.2.6.tar.gz
  • Upload date:
  • Size: 44.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.11.5

File hashes

Hashes for duox-0.2.6.tar.gz
Algorithm Hash digest
SHA256 31cc4a3f5e527ea1dc9c44744fb970f4b155105d373022c4f092c97764124e10
MD5 e1e13bcefc7f0dd8e33669dda9542707
BLAKE2b-256 8b20bd329bde38900e70d8266b6008fc3a27cd977444b559aa992c97065fc0d5

See more details on using hashes here.

File details

Details for the file duox-0.2.6-py3-none-any.whl.

File metadata

  • Download URL: duox-0.2.6-py3-none-any.whl
  • Upload date:
  • Size: 42.9 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.11.5

File hashes

Hashes for duox-0.2.6-py3-none-any.whl
Algorithm Hash digest
SHA256 fea238ef02ed145e4334b0dd87b8f4398f9a8c92ac09858895b920a4cbdef157
MD5 38964ad13f6ee44481663d3f8b2f678b
BLAKE2b-256 ed41d5f1848aadb242a69d131b371ee84a88c5050e74838e7df1aaa92204eded

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page