Frootai

Why FrootAI

The four asymmetries no single-vendor stack can copy:

• 🔀 Vendor-neutral — every play runs identically on Claude, GPT, Gemini, Bedrock, Azure AI Foundry, or your own local Llama. Swap models without rewriting a line. The Terraform of agents.
• 🛡️ Eval-first — every skill and play ships with a regression-tested eval harness; the numbers live in the repo (638/638 spec conformance) and drift breaks CI. Capability is proven, not promised.
• 🌱 Open-source-first — the whole library is MIT and inspectable. No black box, no lock-in; hosted convenience is opt-in, never required.
• ⚡ Edge-native — built on Cloudflare Workers + Durable Objects + R2/KV, so the MCP server runs at the edge with near-zero cold-start, globally.

The Philosophy Behind FrootAI — The Essence of the FAI Engine

FrootAI is an intelligent way of packaging skills, knowledge, and the essential components of the GenAI ecosystem — all synced, not standalone. Infrastructure, platform, and application layers are woven together so that every piece understands and builds on the others. That's what "from the roots to the fruits" means: a fully connected ecosystem where Infra, Platform, and App teams build AI — Frootfully.

FROOT = Foundations · Reasoning · Orchestration · Operations · Transformation

Get Started

⚡ 60-second quickstart — add FrootAI to any MCP-compatible agent (Claude, Copilot, Cursor, Windsurf, Foundry) in one line:

npx frootai-mcp@latest          # zero-install MCP server — 51 tools, vendor-neutral

Point your agent's MCP config at it and you instantly have search, solution plays, eval gates, and knowledge modules. Prefer a different surface? Pick any of the below — each is a published, copy-paste-runnable package:

npx frootai-mcp@latest                          # MCP Server — add to any AI agent
npx frootai init                                 # CLI — scaffold a project
npx skills add frootai/frootai                   # Skills — install FrootAI skills into any agent
npm install frootai                              # npm SDK — import in Node.js/TS
pip install frootai                              # Python SDK
code --install-extension frootai.frootai-vscode  # VS Code Extension
docker run -i ghcr.io/frootai/frootai-mcp        # Docker — zero install

FrootAI's skills follow the Agent Skills open standard and are installable into 70+ agents (GitHub Copilot, Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Goose, and more) via the universal skills installer:

# Interactive picker — choose the skills you need (avoid context rot; install only what's relevant)
npx skills add frootai/frootai

# Install a specific skill into a specific agent, non-interactively
npx skills add frootai/frootai --skill fai-azure-cosmos-modeling --agent github-copilot --yes

# Generate a one-off prompt for a skill without installing it
npx skills use frootai/frootai@fai-mcp-python-generator

Skills are also browsable as themed bundles in Claude Code via /plugin marketplace add frootai/frootai, and render as chips with one-click demo prompts in VS Code / Codex / Cursor.

{
  "mcpServers": {
    "frootai": { "command": "npx", "args": ["frootai-mcp@latest"] }
  }
}

Works with: GitHub Copilot · Claude Desktop · Cursor · Windsurf · Azure AI Foundry · any MCP client

npm SDK — Full Programmatic API

npm install frootai

import { FrootAI, PromptExperiment, PromptVariant, CopilotSession, AgenticLoop, Task } from 'frootai';

const fai = new FrootAI();

// Search, plays, modules
fai.search('RAG architecture');          // BM25-ranked results
fai.plays.get('01');                     // Play details
fai.listModules();                       // 17 FROOT modules (incl. V1 Voice & Speech AI)

// Evaluation quality gates
fai.evaluation.run({ groundedness: 0.95, relevance: 0.88 });

// A/B test prompts
const exp = new PromptExperiment({
  name: 'rag-v2',
  variants: [
    new PromptVariant('control', 'You are a helpful assistant.'),
    new PromptVariant('expert', 'You are an Azure AI expert. Cite sources.'),
  ],
});

// Agentic loop — builder → reviewer → tuner
const loop = new AgenticLoop({ planFile: 'spec/plan.json' });
loop.addTask(new Task('Create RAG pipeline', 'builder'))
    .addTask(new Task('Review security', 'reviewer'));
await loop.run();

The FAI Ecosystem

FAI Protocol — The Binding Glue

FrootAI introduces the FAI Protocol — a context-wiring specification that connects 9 AI primitives (agents, instructions, skills, hooks, workflows, plugins, tools, prompts, guardrails) into evaluated, deployed, production-ready systems.

Primitives:

npm run validate:primitives     # Validate all primitives against schemas
npm run generate:marketplace    # Generate marketplace.json from plugins/
npm run scaffold                # Interactive CLI to create new primitives
node engine/index.js <manifest> # Load a play with the FAI Engine

MCP Server — 51 Tools

Live Eval Scorecard

Every skill and play is gated by a regression eval harness — 638/638 skills pass spec conformance today, across 100+ solution plays and 334 skills. The table below is auto-generated from the eval reports in reports/ and is never hand-edited.

Generated from reports/ by scripts/gen-eval-scorecard.mjs — npm run scorecard. Numbers are read from the eval reports, never hand-edited.

Solution Plays

Every play ships with: fai-manifest.json + .github Agentic OS (agents, instructions, prompts, skills) + DevKit + TuneKit + SpecKit + Bicep infra + evaluation test set

frootai/frootai
├── agents/               238 standalone .agent.md files (WAF-aligned)
├── instructions/         176 standalone .instructions.md files
├── skills/               322 skill folders with SKILL.md
├── hooks/                10 security hooks (secrets, tools, governance, PII, cost...)
├── plugins/              77 themed bundles (1,008 items, 416 categories)
├── workflows/            13 agentic workflows (safe-outputs, NL → YAML)
├── cookbook/              16 step-by-step recipes
├── schemas/              7 JSON schemas validating all primitive types
├── engine/               FAI Engine v0.1 (7 modules — manifest reader → evaluator)
├── solution-plays/       100 deployable plays with DevKit+TuneKit+SpecKit+Bicep
├── npm-mcp/              MCP Server — 51 tools + knowledge.json
├── npm-sdk/              npm SDK & CLI — FrootAI class, 8 commands, 7 modules
├── python-sdk/           Python SDK — offline, zero deps, 7 modules
├── python-mcp/           Python MCP Server
├── vscode-extension/     VS Code extension (4 commands, play browser)
├── functions/            REST API + Agent FAI chatbot
├── docs/                 FROOT knowledge modules (17 modules, ~720 KB)
├── config/               Configurator data + spec templates
├── scripts/              21 build/validate/generate scripts
├── marketplace.json      Auto-generated plugin registry
├── website-data/         8 JSON feeds for frootai.dev
├── workshops/            Hands-on workshops
├── community-plugins/    ServiceNow, Salesforce, SAP
├── bicep-registry/       Azure Bicep modules
├── .github/workflows/    15 CI/CD pipelines
├── .vscode/              15 tasks + schema validation + MCP
├── CONTRIBUTING.md
└── LICENSE (MIT)

GitHub Action

Use frootai/frootai@v6 in any workflow. The v6 action adds MCP federation — pre-attach Tier-1 areas (Azure, Playwright, MS Learn) so evaluations and validations run with full federated context.

Inputs

Outputs

Basic usage (v5 compatible)

- uses: frootai/frootai@v6
  with:
    command: validate

Federation usage

- uses: frootai/frootai@v6
  with:
    command: evaluate
    play: 29-mcp-gateway
    mcp-attach: azure
    version: '6.0.0-alpha.2'

Auto-attach from Play manifest

When a play declares spec/mcp-scope.json with an attached array, the action auto-populates mcp-attach — no workflow change needed:

- uses: frootai/frootai@v6
  with:
    command: evaluate
    play: 29-mcp-gateway       # auto-attaches "azure" from play manifest
    version: '6.0.0-alpha.2'

Sample workflows

• .github/workflows/example-federation.yml — single-area federation with Azure
• .github/workflows/example-multi-mcp.yml — multi-area federation with Azure + Playwright

Secrets handling

The action automatically masks all FROOTAI_* env values in logs via ::add-mask::. For area-specific credentials, use the FROOTAI_SECRET_* naming convention — any env var matching that prefix is masked before the kernel starts:

env:
  FROOTAI_SECRET_AZURE_KEY: ${{ secrets.AZURE_KEY }}

Values injected via ${{ secrets.* }} are also masked by GitHub Actions itself. The double-masking ensures federation debug logs (ACTIONS_STEP_DEBUG=true) never leak credentials.

Links

Contributing

Open source under MIT. See CONTRIBUTING.md.

Star the repo to help others discover FrootAI.