ateam-mcp

Give any AI the ability to build, validate, and deploy production multi-agent systems.

This is an MCP server that connects AI assistants — ChatGPT, Claude, Gemini, Copilot, Cursor, Windsurf, and any MCP-compatible environment — directly to the ADAS platform.

An AI developer says "Build me a customer support system with order tracking and escalation" — and their AI assistant handles the entire lifecycle: reads the spec, builds skill definitions, validates them, deploys to production, and verifies health. No manual JSON authoring, no docs reading, no copy-paste workflows.

Why this matters

Today, building multi-agent systems requires deep platform knowledge, manual configuration, and switching between docs, editors, and dashboards. ateam-mcp eliminates all of that by making the ADAS platform a native capability of the AI tools developers already use.

The AI assistant becomes the developer interface:

Developer: "Create an identity verification agent that checks documents,
            validates faces, and escalates fraud cases"

AI Assistant:
  → reads ADAS spec (adas_get_spec)
  → studies working examples (adas_get_examples)
  → builds skill + solution definitions
  → validates iteratively (adas_validate_skill, adas_validate_solution)
  → deploys to production (adas_deploy_solution)
  → verifies everything is running (adas_get_solution → health)

Developer: "Add a new skill that handles address verification"

AI Assistant:
  → deploys into the existing solution (adas_deploy_skill)
  → redeploys (adas_redeploy)
  → confirms health

No context switching. No manual steps. The full ADAS platform — specs, validation, deployment, monitoring — is available as natural language.

How it reaches the AI community

ChatGPT users

ChatGPT supports MCP connectors in Developer Mode. Users connect by pasting a single URL:

Settings → Connectors → Developer Mode → paste https://mcp.ateam-ai.com

That's it. All 12 ADAS tools appear in ChatGPT. Any ChatGPT Pro, Plus, Business, or Enterprise user can build and deploy multi-agent solutions through conversation.

Claude users

claude.ai and Claude Desktop — add the hosted connector: open Customize > Connectors (Claude Desktop: Customize in the sidebar, then Connectors), click Add custom connector, enter https://mcp.ateam-ai.com, and click Connect. The A-Team sign-in page opens in the browser and asks for your workspace's key: copy it from app.ateam-ai.com/connect (a workspace owner or admin) and paste it there. The key decides the workspace; to switch, Disconnect and Connect again with the other workspace's key.

Claude Code — the hosted connector, signed in the same way:

claude mcp add --transport http ateam https://mcp.ateam-ai.com

then /mcp and authenticate in the browser.

Never paste a key into a chat. ateam_bootstrap's session field says which workspace a session is on and how to switch.

Cursor / Windsurf / VS Code (Copilot)

Add the hosted server to .cursor/mcp.json, mcp_config.json, or .vscode/mcp.json:

{
  "mcpServers": {
    "ateam": { "url": "https://mcp.ateam-ai.com/mcp" }
  }
}

A client with OAuth opens the A-Team sign-in page; one without sends the key as Authorization: Bearer <key> from its own config (see below). A local process (npx -y @ateam-ai/mcp) signs in only through ateam_auth: an ADAS_API_KEY in its environment does not sign it in.

Gemini and other platforms

As MCP adoption grows (it's now governed by the Agentic AI Foundation under the Linux Foundation, co-founded by Anthropic, OpenAI, and Block), every AI platform that implements MCP gets access to ateam-mcp automatically. The remote HTTP endpoint (https://mcp.ateam-ai.com, also at /mcp) works with any client that supports Streamable HTTP transport.

The HTTP endpoint needs a credential on every request: OAuth (clients that support it follow the 401 challenge on their own), or your A-Team API key sent as Authorization: Bearer <key>. An anonymous request is refused with 401; since 0.4.93 there is no anonymous session to call ateam_auth from. See CHANGELOG.md.

Discovery

Developers find ateam-mcp through:

• npm — npm search mcp ai-agents → @ateam-ai/mcp
• Official MCP Registry — registry.modelcontextprotocol.io
• Claude Desktop Extensions — built-in extension browser
• Claude Code Plugin Marketplace — /plugin → Discover tab
• Windsurf MCP Marketplace — built-in marketplace
• VS Code MCP Gallery — Extensions view
• Community directories — Smithery, mcp.so, PulseMCP (30,000+ combined listings)

Available tools

| Tool | What it does | |---|---| | adas_get_spec | Read the ADAS specification — skill schema, solution architecture, enums, agent guides | | adas_get_examples | Get complete working examples — skills, connectors, solutions | | adas_validate_skill | Validate a skill definition through the 5-stage pipeline | | adas_validate_solution | Validate a solution — cross-skill contracts + quality scoring | | adas_deploy_solution | Deploy a complete solution to production | | adas_deploy_skill | Add a skill to an existing solution | | adas_deploy_connector | Deploy a connector to ADAS Core | | adas_list_solutions | List all deployed solutions | | adas_get_solution | Inspect a solution — definition, skills, health, status, export | | adas_update | Update a solution or skill incrementally (PATCH) | | adas_redeploy | Push changes live — regenerates MCP servers, deploys to ADAS Core | | adas_solution_chat | Talk to the Solution Bot for guided modifications | | ateam_workspace_open | For a chat agent with no shell: open one private workspace (files in /workspace) for the conversation | | ateam_workspace_exec | Run one command in the workspace (100 s foreground; run_in_background for long builds) | | ateam_workspace_status | Check the workspace: unsaved changes, a long save or background command | | ateam_workspace_save | Make the workspace's work live (the only way the other tools see it) | | ateam_workspace_close | Finish; optionally save first. Brings back the single-file write tools withheld while it was open |

The ateam_workspace_* tools are listed whenever you are signed in. A workspace the platform cannot give (a trial workspace, a key tied to no person) is refused by the call itself, with the single-file tools to use instead.

Setup

# Clone
git clone https://github.com/ariekogan/ateam-mcp.git
cd ateam-mcp

# Install
npm install

# Run (a local stdio process)
npm start

A local process signs in with ateam_auth, given a workspace key the agent reads from a file outside the chat; an ADAS_API_KEY in the environment does not sign it in and is never sent. ADAS_API_URL in its environment points it at a self-hosted API. Most people use the hosted connector instead (above).

Architecture

┌─────────────────────────────────────────────┐
│  AI Environment                             │
│  (ChatGPT / Claude / Cursor / Windsurf)     │
│                                             │
│  Developer: "build me a support system"     │
└──────────────────┬──────────────────────────┘
                   │ MCP protocol
                   │ (stdio or HTTP)
┌──────────────────▼──────────────────────────┐
│  ateam-mcp                                  │
│  12 tools — spec, validate, deploy, manage  │
└──────────────────┬──────────────────────────┘
                   │ HTTPS
                   │ X-ADAS-TENANT / X-API-KEY
┌──────────────────▼──────────────────────────┐
│  ADAS External Agent API                    │
│  api.ateam-ai.com                           │
└──────────────────┬──────────────────────────┘
                   │
┌──────────────────▼──────────────────────────┐
│  ADAS Core                                  │
│  Multi-agent runtime                        │
└─────────────────────────────────────────────┘

License

MIT