GitHub - drivelineresearch/moa-x: Mixture-of-Agents-X - utilizing the MoA framework in agentic coding tools.

A small, CLI-native take on the 2024 Mixture-of-Agents paper, pointed at a different job: producing repo-grounded implementation plans for coding agents instead of chat answers. The default roster puts proposers from three model labs to work — Google agy-gemini-pro (gemini-3.1-pro-high), xAI grok (opencode-go/grok-4.5), and OpenAI codex-luna (gpt-5.6-luna) — reading the repo in parallel, doing their own web research, and each writing an independent plan. Three refiners—Alibaba Qwen qwen3.8-max-preview, DeepSeek V4 Pro, and Anthropic claude-opus-5—then refine in broadcast mode (every refiner sees every plan). The shipped defaults and all recommended Web UI modes run Layer 3 through the recorded Codex path with gpt-5.6-sol at xhigh reasoning. Claude Opus 5 is also available as an aggregator, while Fable 5 is available only as a warning-gated, quota-heavy aggregator option.

Built to run inside Claude Code as a skill, or from the local Web UI. Standalone Python works too. The harness ships curated routes across four execution harnesses (codex, claude, opencode, agy) and the roster — which models run at which layer, and how many — is pure config. API-based auth and more providers are already supported. See "Contributions we'd prioritize" below for the remaining gaps.

Qwen Cloud Token Plan powers the default qwen refiner (qwen-token-plan/qwen3.8-max-preview through OpenCode). Its dedicated sk-sp-... key stays in .env; see docs/config.md.

TL;DR

# 1. Install the CLIs (see docs/install.md for details)
npm i -g @openai/codex               && codex login
curl -fsSL https://opencode.ai/install | bash   # then: opencode auth login,
                                                 # or export MOONSHOT_API_KEY
# claude CLI: https://docs.claude.com/en/docs/claude-code/quickstart
# AGY/Antigravity: install/update Antigravity, run `agy install`, then sign in

# 2. Install as a Claude Code skill
cp -r harness ~/.claude/skills/mixture-of-agents

# 3. Inside Claude Code, in any project
/mixture-of-agents

Local Web UI

The Flask control room queues runs, streams phase events, shows local provider health, and indexes current and historical .moa sessions in SQLite. It uses the same HOME, PATH, environment, and authenticated CLI accounts as the server process; credentials are never copied into the browser or database. Tasks can run without a repository and may include up to 10 durable local file uploads (25 MB each). PDF, Markdown, common text formats, and locally OCRed images are copied into an isolated run snapshot, converted into one bounded Markdown context packet, and embedded in every proposer, refiner, and aggregator prompt. Local workspaces are optional. When repository context is useful, the GitHub picker can shallow-clone only an explicitly allowed GitHub owner through the machine's authenticated gh CLI. Browser-generated profile IDs remain in localStorage; profile names/settings, jobs, events, and history are retained in SQLite without adding a login system.

git clone https://github.com/drivelineresearch/moa-x.git
cd moa-x
python3 -m venv .venv
.venv/bin/pip install -r requirements-web.txt
MOA_WEBUI_GITHUB_OWNER=your-github-user-or-org \
  .venv/bin/python -m harness.webui

Open http://localhost:7340. There is no frontend compilation step: the versioned HTML, CSS, JavaScript, and image assets are served directly by Flask/Waitress. The safe default bind is 127.0.0.1. To opt into trusted-LAN access, set MOA_WEBUI_HOST=0.0.0.0; because the UI intentionally has no login, never expose that listener directly to the public internet. Set MOA_WEBUI_PORT to change the port and MOA_WEBUI_WORKSPACE_ROOTS (colon-separated on Linux/macOS) to bound backend path operations such as history import. New runs use Task only or an allowlisted GitHub repository. See docs/webui.md for clean-clone setup, Task-only storage, GitHub configuration, development, and the HTTP surface.

AGY is the curated Google route and reuses the account already signed into agy; its live model probe determines which Gemini routes are available.

Architecture at a glance

Layer 0 — Scout brief           (parent Claude, in-place)
Layer 1 — Proposers (parallel)    default: Gemini Pro + Grok + GPT-5.6 Luna
Layer 2 — Broadcast refiners      default: Qwen + DeepSeek + Opus, each sees ALL proposals
Layer 3 — Aggregator              default: recorded GPT-5.6 Sol at xhigh

The roster is config-driven; every named route in the default pipeline comes from a deliberately named lab: Google, xAI, OpenAI, Alibaba, DeepSeek, and Anthropic. The UI groups and illustrates routes by model lab, independently of the CLI used to execute them.

Every run also writes a self-contained .moa/<session>/report.html — a zero-network visual post-mortem (3D pipeline, per-agent Gantt, proposer plans, refiner verdict matrix, an evidence-weighted living decision map, interactive final-step lineage, aggregated plan, and raw logs). Open it in a browser; details in docs/report.md.

Typical wall-clock is roughly 12–25 minutes for research-heavy work, with provider latency determining the tail. Use it for non-trivial architecture work, not one-line fixes. Background in docs/architecture.md.

Docs

docs/install.md: install the CLIs, verify, install as a Claude Code skill
docs/usage.md: running via /mixture-of-agents (primary) or standalone
docs/webui.md: local control-room setup, persistence, uploads, GitHub workspaces, security
docs/config.md: .env + harness/config.yaml, MOA_* knob table, precedence, roster swaps
docs/architecture.md: the four layers, why broadcast, why this roster
docs/report.md: the self-contained HTML run report (report.html) — timeline, verdicts, evidence-weighted decision map, and exact lineage
docs/assets.md: asset provenance, font policy, animation sources, and contribution rules
CONTRIBUTING.md: dev setup, PR protocol, where help is welcome
SECURITY.md: private vulnerability reports
CLAUDE.md / AGENTS.md: guidance for coding agents working on this repo (AGENTS.md points at CLAUDE.md)

Repo layout

README.md              this file
CLAUDE.md              agent guidance for this repo
AGENTS.md              pointer to CLAUDE.md for coding agents
CONTRIBUTING.md        contributor guide
CHANGELOG.md           release notes
SECURITY.md            vulnerability reporting
LICENSE                MIT
.env.example           copy to .env to override harness defaults
docs/                  longer-form docs by topic (+ brand images)
harness/               orchestrator, adapters, prompts, schemas
  SKILL.md             Claude Code skill manifest
  README.md            skill-internal notes (lives with harness/ when copied into ~/.claude/skills/)
  config.example.yaml  copy to harness/config.yaml to override defaults
  prompts/             scout / proposer / refiner / aggregator
  report/              HTML report template + embedded illustration assets
  scripts/             orchestrator + adapters + deterministic decision map + report + tests
  webui/               Flask control plane, SQLite store, worker, and frontend
requirements-cli.txt   install/auth notes for the provider CLIs
requirements-web.txt   optional Flask control-room dependencies

Contributions we'd prioritize

The core roster, named-provider system, API-key auth paths, Qwen Token Plan, phase checkpoints, and HTML reporting are now shipped. The highest-leverage remaining contributions are:

A CLI-only raw-spec convenience command. The Web UI can create a job from a raw goal and drive every recorded phase. run_moa.py also handles the proposer/refiner layers and recorded Layer 3 from a shell. The remaining gap is a single non-Web CLI command that creates Layer 0 and drives all phases without a parent session.
Usage, quota, and cost observability. Capture the token/usage metadata each CLI exposes, normalize it into the manifest and HTML report, distinguish subscription from metered runs, and make unknown cost explicit. A safe budget control could stop later dispatches before a configured ceiling is exceeded; it must not pretend it can undo an already-billed request.
Tested provider recipes, not just model-name examples. Qwen Token Plan, OpenCode Go DeepSeek, GLM, and Grok routes are already built in. Contributions for MiniMax, Mistral, or another credible coding model should include a reproducible config, credential preflight, captured parser fixtures, and an end-to-end smoke-test result. Most should use the existing OpenCode adapter; discuss a genuinely new harness in an issue first.
CLI compatibility and recovery hardening. Add version/capability probes, fixture-based coverage for real failure envelopes, clearer auth/quota/model diagnostics, and resumable recovery paths that avoid rerunning successful agents after an interrupted session.

API-key billing itself is no longer a missing feature: Codex supports API-key login, Claude accepts ANTHROPIC_API_KEY, and OpenCode routes provider keys. The missing layer is normalized usage and cost telemetry across those different billing modes.

See CONTRIBUTING.md for the PR protocol.

Status

Active reference implementation, currently v0.4.1. The default six-lab roster and Qwen Token Plan route have been exercised end to end; offline CI covers configuration, schemas, adapters, checkpoint recovery, recorded Layer 3, and self-contained HTML report generation. Contributions are welcome; see CONTRIBUTING.md, and release notes are in CHANGELOG.md. Security reports go through SECURITY.md.

License

Author

Kyle Boddy.