Skip to content

Getting Started

Connect your AI assistant to ACCESS-CI services and start exploring cyberinfrastructure resources. These servers use the Model Context Protocol (MCP), an open standard supported by Claude and other AI tools.

What You Get

Once connected, you can ask your AI questions like:

  • "What GPU resources are available on ACCESS-CI?"
  • "Are there any current outages or upcoming maintenance on Delta?"
  • "Show me upcoming workshops on parallel computing"
  • "What machine learning software is available on Anvil?"
  • "What NSF-funded projects are doing work in quantum computing?"

Your AI queries the relevant ACCESS services and gives you a synthesized answer — no need to navigate multiple websites or APIs.

Logging In

Most tools work without logging in. Browsing, searching, and reading data are all available anonymously. You only need to authenticate if you want to create or manage content — for example, drafting an announcement or viewing your own draft events.

Which servers require login?

ServerAnonymous useRequires login
Compute ResourcesAll tools
System StatusAll tools
Software DiscoveryAll tools
XDMoDAll tools
AllocationsAll tools
NSF AwardsAll tools
EventsAll tools
Announcementssearch_announcements, suggest_tags, suggest_summary, get_announcement_contextcreate_announcement, update_announcement, delete_announcement, get_my_announcements
Affinity GroupsAll tools

How to log in

When you connect a server that supports authentication (currently Announcements), Claude will prompt you to authorize via CILogon — the same single sign-on system used by other ACCESS services.

  1. After adding the connector, Claude will show a CILogon authorization page in your browser
  2. Choose your home institution from the list and sign in with your institutional credentials
  3. Approve the authorization
  4. Return to Claude — the connector is now authenticated

You can skip this step if you only need read-only access. The server will still work for anonymous tools; you just won't be able to use the management tools.

Logging out / revoking access

To revoke access for a connector, remove it in Customize > Connectors in Claude. The next time you add it, you'll be prompted to log in again.

Available Servers

All servers are hosted at https://mcp.access-ci.org. You can add as many or as few as you need — start with one and add more later.

ServerURLDescription
Compute Resourceshttps://mcp.access-ci.org/compute-resources/sseHardware specs, GPU availability, system capabilities
System Statushttps://mcp.access-ci.org/system-status/sseCurrent outages, maintenance schedules, incidents
Software Discoveryhttps://mcp.access-ci.org/software-discovery/sseSoftware packages available across ACCESS resources
XDMoDhttps://mcp.access-ci.org/xdmod/sseUsage statistics, visualizations, resource utilization
Allocationshttps://mcp.access-ci.org/allocations/sseResearch projects and allocation trends
NSF Awardshttps://mcp.access-ci.org/nsf-awards/sseNSF funding data and cross-referencing
Eventshttps://mcp.access-ci.org/events/sseWorkshops, webinars, training sessions
Announcementshttps://mcp.access-ci.org/announcements/sseCommunity news and updates
Affinity Groupshttps://mcp.access-ci.org/affinity-groups/sseCommunity groups, events, and knowledge base

For XDMoD Data Analytics, see the separate setup section below — it requires a personal API token and works only via Claude Code, not as a Claude.ai connector.

Quick Start

Claude Desktop or claude.ai

Add ACCESS servers as connectors. Setup takes about a minute per server.

  1. In Claude, open Customize (palette/settings icon) > Connectors
  2. Click "Add custom connector"
  3. Paste https://mcp.access-ci.org/compute-resources/sse into the Server URL box
  4. Save and authorize when prompted
  5. Repeat for any other servers from the table above

TIP

You don't need to add all servers at once. Start with one or two that match your needs, then add more later.

Try it: Open a chat and ask "What GPU resources are available on ACCESS-CI?"

Claude Code (CLI)

Requires Node.js 20+. Add all servers with:

bash
claude mcp add access-compute-resources -s user -- npx mcp-remote https://mcp.access-ci.org/compute-resources/sse
claude mcp add access-system-status -s user -- npx mcp-remote https://mcp.access-ci.org/system-status/sse
claude mcp add access-software-discovery -s user -- npx mcp-remote https://mcp.access-ci.org/software-discovery/sse
claude mcp add access-xdmod -s user -- npx mcp-remote https://mcp.access-ci.org/xdmod/sse
claude mcp add access-allocations -s user -- npx mcp-remote https://mcp.access-ci.org/allocations/sse
claude mcp add access-nsf-awards -s user -- npx mcp-remote https://mcp.access-ci.org/nsf-awards/sse
claude mcp add access-events -s user -- npx mcp-remote https://mcp.access-ci.org/events/sse
claude mcp add access-announcements -s user -- npx mcp-remote https://mcp.access-ci.org/announcements/sse
claude mcp add access-affinity-groups -s user -- npx mcp-remote https://mcp.access-ci.org/affinity-groups/sse

Restart Claude Code after adding.

Other MCP Clients

The Model Context Protocol is supported by a growing number of AI tools beyond Claude. The ACCESS MCP servers should work with any client that supports remote MCP servers over SSE or Streamable HTTP transport.

The connection URL pattern is the same for all clients:

https://mcp.access-ci.org/<server-name>/sse

Setup instructions vary by client — refer to each tool's documentation for adding custom MCP servers. Some clients only support local (stdio) MCP servers; for those, you can use npx mcp-remote <url> as a stdio-to-SSE bridge (the same approach used by Claude Code above).

We've focused our testing on Claude. If you successfully connect another client — or run into trouble — we'd love to hear about it. Open a support ticket so we can improve the docs or address compatibility gaps.

Try These Queries

Once connected, ask your AI questions like:

Resources & hardware:

  • "What GPU resources are available on ACCESS-CI?"
  • "Show me the hardware specs for Delta"
  • "Which systems support NVIDIA A100 GPUs?"

System status:

  • "Are there any current outages?"
  • "What maintenance is scheduled this week?"

Software:

  • "Is TensorFlow available on Anvil?"
  • "What Python ML libraries are installed on DeltaAI?"

Events & training:

  • "Show me upcoming workshops on parallel computing"
  • "What training events are happening this month?"

Allocations & research:

  • "What research projects use GPU resources?"
  • "Find allocations related to climate modeling"

NSF funding:

  • "What NSF-funded projects are doing work in quantum computing?"
  • "Show NSF awards for the University of Michigan"

Optional: XDMoD Data Analytics

The XDMoD Data server provides deep analytics across jobs, allocations, cloud usage, resource specifications, and more using the XDMoD data analytics framework. It requires a personal API token.

Compatibility note

The XDMoD Data server only works with Claude Code or other clients that can pass HTTP headers (via the mcp-remote --header flag or stdio). It is not compatible with Claude.ai connectors, because the SSE transport used by connectors cannot pass per-request authentication headers.

Get your API token:

  1. Sign in to xdmod.access-ci.org with your ACCESS credentials
  2. Click My Profile in the top-right corner
  3. Click the API Token tab
  4. Generate and copy your token

Add to Claude Code:

bash
claude mcp add access-xdmod-data -s user -- npx mcp-remote https://mcp.access-ci.org/xdmod-data/sse --header "X-XDMoD-Token:your-token-here"

Add to other clients with header support — example config:

json
"access-xdmod-data": {
  "command": "npx",
  "args": ["mcp-remote", "https://mcp.access-ci.org/xdmod-data/sse", "--header", "X-XDMoD-Token:your-token-here"]
}

Troubleshooting

"Server disconnected" Error

  • Verify the hosted servers are accessible: curl https://mcp.access-ci.org/compute-resources/health
  • Check your internet connection
  • Restart Claude Desktop (fully quit with Cmd+Q, then reopen)

npm/npx Not Found (for Claude Code)

Install Node.js LTS which includes npm and npx.

macOS (Homebrew):

bash
brew install node

Windows:

powershell
winget install OpenJS.NodeJS.LTS

Tool Schemas Look Outdated

MCP clients cache tool definitions when they connect. After a server update, you may need to reconnect:

  • Claude Desktop: Fully quit (Cmd+Q on macOS) and reopen
  • Claude Code: Restart the CLI

Connectors Not Showing Up

  1. Verify the URL is exactly right (including the /sse suffix)
  2. Check that the connector was authorized after adding
  3. Restart Claude completely

Next Steps


For Developers

Want to contribute or extend the MCP servers? Set up a local development environment.

Clone and Build

bash
git clone https://github.com/necyberteam/access-mcp.git
cd access-mcp
npm install        # Install all workspace dependencies
npm run build      # Build all packages

Note: The xdmod-mcp-data package requires Python 3.11+. See its README for setup instructions.

Run Tests

bash
npm test                  # Unit tests
npm run test:integration  # Integration tests (requires API access)
npm run test:all          # All tests

Test Local Changes

Run a server locally and connect Claude Desktop to it:

bash
# Start a server in HTTP mode
PORT=3002 node packages/compute-resources/dist/index.js

Then configure Claude Desktop to use your local server:

json
{
  "mcpServers": {
    "access-compute-resources-dev": {
      "command": "npx",
      "args": ["mcp-remote", "http://localhost:3002/sse"]
    }
  }
}

Self-Hosting (Docker)

Organizations can run their own instance of the MCP servers using Docker.

Using Pre-built Images

Images are published to GitHub Container Registry on every push to main.

  1. Create deployment directory:

    bash
    mkdir ~/access-mcp && cd ~/access-mcp
  2. Download compose file:

    bash
    curl -o docker-compose.yml https://raw.githubusercontent.com/necyberteam/access-mcp/main/docker-compose.prod.yml
  3. Create environment file:

    bash
    cat > .env << 'EOF'
    GITHUB_REPOSITORY=necyberteam/access-mcp
    SDS_API_KEY=your-sds-api-key-here
    XDMOD_API_TOKEN=your-xdmod-api-token-here
    ACCESS_MCP_SERVICES=nsf-awards=http://mcp-nsf-awards:3000
    EOF
  4. Start services:

    bash
    docker login ghcr.io -u YOUR_GITHUB_USERNAME
    docker compose pull
    docker compose up -d

Connect to Your Deployment

Update the Claude Desktop config to point to your server:

json
{
  "mcpServers": {
    "access-compute-resources": {
      "command": "npx",
      "args": ["mcp-remote", "http://YOUR_SERVER_IP:3002/sse"]
    }
  }
}

Replace YOUR_SERVER_IP with your server's hostname or IP. Each MCP server runs on its own port (3002–3012); see docker-compose.prod.yml for the full mapping.

Environment Variables

VariableRequiredDescription
GITHUB_REPOSITORYYesRepository path for image pulls
SDS_API_KEYFor software-discoverySoftware Discovery Service API key
XDMOD_API_TOKENFor xdmod-dataXDMoD API token (see XDMoD Data server docs for setup and per-user token plans)
ACCESS_MCP_SERVICESFor NSF integrationInter-service communication endpoints

Automated Deployment

The repository includes GitHub Actions for CI/CD. Configure these secrets for automated deployment:

  • PRODUCTION_HOST - Server hostname or IP
  • PRODUCTION_SSH_KEY - SSH private key for deployment

See .github/workflows/deploy-production.yml for details.