MultAI

Submit a research prompt to 7 AI platforms simultaneously — all from Claude Code.

MultAI is a Claude Cowork/Code plugin that automates multi-AI research workflows. It uses Playwright CDP automation to submit prompts to Claude.ai, ChatGPT, Microsoft Copilot, Perplexity, Grok, DeepSeek, and Google Gemini in parallel, extracts their responses, and synthesizes the results into structured deliverables — market landscape reports, solution capability matrices, comparison spreadsheets, and more.

---

What It Does

You (Claude Code)
      │
      ▼
┌──────────────────────────────────────────────────────┐
│  Orchestrator  (intelligent router — Phase 0)         │
│                                                        │
│  ┌──────────────────┐   ┌────────────────────────┐   │
│  │ landscape-       │   │ solution-researcher    │   │
│  │ researcher       │   │ product deep-dive      │   │
│  └──────────────────┘   └────────────────────────┘   │
│  ┌──────────────────┐   ┌────────────────────────┐   │
│  │ comparator       │   │ direct multi-AI        │   │
│  │ XLSX matrix ops  │   │ + generic consolidator │   │
│  └──────────────────┘   └────────────────────────┘   │
└──────────────────────────────────────────────────────┘
      │
      ▼
Playwright Engine ──► 7 AI Platforms ──► reports/

Capability	Detail
Parallel submission	All 7 platforms run concurrently via asyncio.gather()
Intelligent routing	Orchestrator analyzes intent and dispatches to the right skill automatically
Market landscape reports	9-section structured reports (Top 20 commercial + OSS, positioning matrices, trends)
Solution research	Deep-dive on a product URL — capability inventory, competitive context, XLSX scoring
XLSX comparison matrix	Capability matrix auto-scored and reordered across platforms
Rate limiting	Per-platform budget + cooldown + daily cap, persisted across sessions
Agent fallback	Vision-based browser-use agent kicks in automatically when a UI selector fails
DEEP mode	Activates Deep Research / Research mode on each platform (where available)
Prompt echo filtering	Automatically strips platform echoes from extracted responses
Report viewer	preview.html?report=<path> renders any Markdown report with charts
Self-improving skills	Each skill appends a run log and can update its own templates after every run
Shared domain knowledge	domains/{domain}.md is enriched by both research skills after every run

---

Supported Platforms

Platform	URL	Notes
Claude.ai	https://claude.ai	Pro plan recommended for DEEP mode
ChatGPT	https://chat.openai.com	Plus plan for Deep Research
Microsoft Copilot	https://copilot.microsoft.com	Free tier works
Perplexity	https://www.perplexity.ai	Pro for Deep Research
Grok	https://grok.com	X/Twitter account required
DeepSeek	https://chat.deepseek.com	Free tier works
Google Gemini	https://gemini.google.com	Google account required

---

Quick Start

1 — Prerequisites

• Claude Code v1.0.33 or later — run claude --version to check; update with brew upgrade claude-code or npm update -g @anthropic-ai/claude-code
• Python 3.11+, Google Chrome installed

2 — Install via Claude Code Plugin Manager

# Step 1 — register the repo as a marketplace (one-time):
/plugin marketplace add alo-exp/multai

# Step 2 — install the plugin from that marketplace:
/plugin install multai@multai

Skills are namespaced: /multai:orchestrator, /multai:solution-researcher, etc. Run /reload-plugins if skills don't appear immediately.

Python dependencies (playwright, openpyxl, Chromium browser) are installed automatically on the first session start via a SessionStart hook (hook → install.sh → setup.sh). No manual setup required.

Agent fallback (optional): To enable the vision-based browser-use fallback, run manually:

bash "$(find ~/.claude/plugins/cache -name setup.sh | head -1)" --with-fallback

Alternative — Local / Dev Install

git clone https://github.com/alo-exp/multai.git
cd multai
bash setup.sh            # creates .venv, installs pip deps + Playwright Chromium + .env template
# optional agent fallback:
bash setup.sh --with-fallback

# Load directly without marketplace registration:
claude --plugin-dir ./multai

3 — Log in to platforms

Open Chrome and log in to each of the 7 AI platforms above. The engine re-uses your existing Chrome profile — no credentials are stored by this tool.

4 — Set optional API keys

# ~/.zshrc or ~/.bashrc
export GOOGLE_API_KEY="..."      # free from aistudio.google.com — enables Gemini agent fallback
export ANTHROPIC_API_KEY="..."   # from console.anthropic.com — enables Claude agent fallback

4 — Open your project in Claude Code and use a skill

# Route a landscape research request:
"Run a market landscape analysis on DevOps platforms for SMBs"

# Route a product deep-dive:
"Research humanitec.com"

# Route a matrix operation:
"Add Harness to the comparison matrix"

# Direct multi-AI (no specific route):
"What are the main trade-offs between Rust and Go for backend services?"

The orchestrator announces its routing decision before proceeding — you can override it.

---

Skills

orchestrator — Intelligent Router + Engine Owner

The entry point for all workflows. Phase 0 analyzes your intent and routes to the correct specialist skill. For direct multi-AI queries it owns the full pipeline itself.

Triggers: any research or comparison prompt; explicitly routing to another skill

Routing rules:

"landscape" / "market map" / "ecosystem" / "vendor landscape"
  → landscape-researcher

Product URL or named product + research intent
  → solution-researcher

"comparison matrix" / "add platform" / "update matrix"
  → comparator

Everything else
  → direct multi-AI (orchestrator handles end-to-end)

---

landscape-researcher — Market Landscape Reports

Produces a 9-section structured Market Landscape Report across commercial and OSS solutions for any product category.

Triggers: "landscape analysis on X", "market map for X", "competitive landscape", "ecosystem overview"

Output: reports/{task-name}/{Category} - Market Landscape Report.md + auto-launched browser preview

Report sections:

1. Market Definition & Scope
2. Market Overview (size, CAGR, drivers)
3. Competitive Positioning (2×2, Wave-style, Value Curve)
4. Key Industry Trends (5–10)
5. Top 20 Commercial Solutions for SMBs
6. Top 20 OSS Solutions
7. Buying Guidance & Shortlist Profiles
8. Future Outlook & Emerging Disruptors
9. Source Reliability Assessment

---

solution-researcher — Product Deep-Dives

Researches a specific product or URL in depth — capabilities, integrations, pricing, competitive context — and optionally scores it in the comparison matrix.

Triggers: "research humanitec.com", "evaluate Backstage", "analyze Port.io"

Output: reports/{task-name}/{Product} - Consolidated Intelligence Report.md + optional XLSX matrix scoring

---

comparator — XLSX Capability Matrix

Maintains and operates the capability comparison matrix spreadsheet.

Triggers: "add X to the matrix", "update comparison matrix", "score Harness", "combo column", "verify ticks"

Scripts:

• skills/comparator/matrix_ops.py — add platforms, update scores, apply combos
• skills/comparator/matrix_builder.py — rebuild matrix from raw response archives

---

consolidator — Response Synthesis

Synthesizes raw multi-AI responses into a single structured report. Called automatically by other skills; rarely invoked directly.

With a consolidation guide: the guide is the sole structural authority — used by landscape-researcher (9-section guide) and solution-researcher (CIR guide).

Without a guide: produces a well-structured synthesis using its own judgement.

---

Project Structure

multai/
├── .claude-plugin/
│   ├── plugin.json           ← Claude Code Plugin manifest
│   └── hooks.json            ← SessionStart hook (runs install.sh → setup.sh once)
├── skills/
│   ├── orchestrator/
│   │   ├── SKILL.md          ← Router + phases
│   │   ├── platform-setup.md ← Per-platform login/injection notes
│   │   └── engine/           ← Playwright automation engine
│   │       ├── orchestrator.py
│   │       ├── config.py
│   │       ├── rate_limiter.py
│   │       ├── agent_fallback.py
│   │       ├── prompt_echo.py
│   │       ├── collate_responses.py
│   │       ├── utils.py
│   │       └── platforms/    ← claude_ai.py chatgpt.py copilot.py …
│   ├── consolidator/
│   │   └── SKILL.md
│   ├── landscape-researcher/
│   │   ├── SKILL.md
│   │   ├── prompt-template.md
│   │   ├── consolidation-guide.md
│   │   └── launch_report.py  ← stdlib-only HTTP server + browser open
│   ├── solution-researcher/
│   │   ├── SKILL.md
│   │   ├── prompt-template.md
│   │   └── consolidation-guide.md
│   └── comparator/
│       ├── SKILL.md
│       ├── matrix_ops.py
│       └── matrix_builder.py
├── domains/
│   └── devops-platforms.md   ← Shared domain knowledge (enriched per run)
├── reports/
│   └── preview.html          ← Query-param driven report viewer
├── docs/
│   ├── Architecture-and-Design.md
│   ├── SRS.md
│   ├── Test-Strategy-and-Plan.md
│   └── CICD-Strategy-and-Plan.md
├── tests/                    ← pytest suite
├── setup.sh                  ← Canonical bootstrap — creates .venv, installs deps
├── install.sh                ← Plugin hook delegate → setup.sh
├── pyproject.toml            ← Python packaging spec
├── settings.json             ← Default Claude Cowork/Code plugin permissions
├── USER-GUIDE.md             ← End-user guide (non-technical)
└── CONTRIBUTOR-GUIDE.md      ← Technical contributor reference

---

Engine CLI Reference

The orchestration engine can also be driven directly from the terminal:

# Check platform budgets (no browser opened):
python3 skills/orchestrator/engine/orchestrator.py --budget --tier free

# Submit a prompt in REGULAR mode:
python3 skills/orchestrator/engine/orchestrator.py \
  --prompt "What is the CAP theorem?" \
  --mode REGULAR \
  --task-name my-research

# Submit from a file in DEEP mode:
python3 skills/orchestrator/engine/orchestrator.py \
  --prompt-file /tmp/my-prompt.md \
  --mode DEEP \
  --task-name deep-research-20260316

# Collate archived responses into a single file:
python3 skills/orchestrator/engine/collate_responses.py \
  --archive-dir reports/my-research/

Output is written to reports/{task-name}/ — one .md file per platform, plus a status.json summary.

---

Rate Limiting

Each platform has a per-session budget, a post-use cooldown, and a daily cap:

Tier	Budget	Cooldown	Daily Cap
free	3 requests	30 min	5
pro	10 requests	5 min	20

State is persisted in .rate-limit-state.json (gitignored) so limits carry across sessions. The engine prints remaining budget before each run and skips platforms in cooldown.

---

Agent Fallback

When a Playwright selector fails (e.g. platform UI changed), a browser-use vision agent takes over:

1. Anthropic key present → Claude Sonnet is the agent LLM
2. Google key present → Gemini 2.0 Flash is the agent LLM (free tier at aistudio.google.com)
3. Neither key → fallback silently disabled; Playwright exception propagates

# Set in your shell profile:
export GOOGLE_API_KEY="AIza..."        # recommended (free)
export ANTHROPIC_API_KEY="sk-ant-..."  # alternative

The fallback venv is set up by setup.sh --with-fallback and auto-detected by the engine.

---

Report Viewer

Reports are rendered by a self-contained preview.html served locally:

# From the landscape-researcher skill (auto-launched):
python3 skills/landscape-researcher/launch_report.py \
  --report-dir market-landscape-20260316-1430 \
  --report-file "DevOps Platforms - Market Landscape Report.md" \
  --port 7788

Or open manually:

http://localhost:7788/preview.html?report=market-landscape-20260316-1430/DevOps%20Platforms%20-%20Market%20Landscape%20Report.md

The viewer renders Markdown, highlights code blocks, and injects positioning charts when the report contains a matrix section.

---

Domain Knowledge

domains/{domain}.md is a shared living document enriched after every research run:

• landscape-researcher adds: new vendor archetypes, market-wide trend signals, category boundary changes, emerging vendors
• solution-researcher adds: new capability categories, terminology introduced by a vendor, feature-name equivalences

Both skills propose additions and ask for approval before writing. Additions are always append-only and timestamped.

---

Running Tests

# Full test suite (using the project venv):
skills/orchestrator/engine/.venv/bin/python -m pytest tests/ -v

# Alternatively, from the engine directory:
cd skills/orchestrator/engine && .venv/bin/python -m pytest ../../../tests/ -v

# Specific test file:
skills/orchestrator/engine/.venv/bin/python -m pytest tests/test_rate_limiter.py -v

# Engine budget check (smoke test — no browser):
python3 skills/orchestrator/engine/orchestrator.py --budget --tier free

---

Documentation

Document	Description
USER-GUIDE.md	End-user guide — installation, skills overview, viewing reports
CONTRIBUTOR-GUIDE.md	Contributor reference — CLI flags, platform internals, tests, CI/CD
docs/Architecture-and-Design.md	System topology, data flows, design decisions
docs/SRS.md	Software Requirements Specification
docs/Test-Strategy-and-Plan.md	Test cases, coverage strategy
docs/CICD-Strategy-and-Plan.md	CI/CD pipeline, make targets, deployment
CHANGELOG.md	Version history

---

Requirements

Requirement	Version	Notes
Python	≥ 3.11	3.13 required for agent fallback
Google Chrome	latest	Must be installed; engine re-uses your profile
playwright	≥ 1.40.0	Installed by setup.sh
openpyxl	≥ 3.1.0	Installed by setup.sh
Claude Code	latest	Skills are invoked as Claude Cowork/Code plugin skills

---

License

MIT — see LICENSE or the license field in pyproject.toml.

---

Author

alo-exp · GitHub · User Guide · Contributor Guide