MCP — agent access

One URL — https://data.sierragridteam.org/mcp — speaking MCP over Streamable HTTP. Clients POST JSON-RPC 2.0; there's no server-initiated stream, so a GET returns 405. It's a thin adapter over the /api/v1 API: geometry is stripped for token efficiency, and the fail-loud honesty contract is preserved on every result.

Add it as a remote MCP server. Clients that speak Streamable HTTP directly take the URL:

{
  "mcpServers": {
    "sierra-grid": {
      "url": "https://data.sierragridteam.org/mcp"
    }
  }
}

For a client that only supports stdio servers, bridge it with mcp-remote:

npx mcp-remote https://data.sierragridteam.org/mcp

Because it's keyless and read-only, there's nothing to authorize — you just point the client at the URL and go.

Claude

Claude.ai / Claude Desktop. On a paid plan, open Settings → Connectors → Add custom connector, give it a name (Sierra Grid), and paste the endpoint URL:

https://data.sierragridteam.org/mcp

It connects with no OAuth step — the eight tools appear in the composer's tool menu, and you can ask questions like the ones below in plain language.

Claude Code. Add it as a remote HTTP server from the CLI:

claude mcp add --transport http sierra-grid https://data.sierragridteam.org/mcp

OpenAI / ChatGPT

ChatGPT. On Pro, Business, or Enterprise, enable Settings → Connectors (developer mode), add a custom MCP connector, and paste the same https://data.sierragridteam.org/mcp URL — no authentication.

OpenAI API. The Responses API takes a remote MCP server as a hosted mcp tool — the model calls the Grid's tools directly, server-side:

# OpenAI Responses API — the endpoint as a hosted MCP tool
client.responses.create(
    model="gpt-5",
    tools=[{
        "type": "mcp",
        "server_label": "sierra_grid",
        "server_url": "https://data.sierragridteam.org/mcp",
        "require_approval": "never",
    }],
    input="Is it safe near Arnold, CA right now?",
)

Every tool accepts a flexible location — a place slug (ebbetts-pass), a street address, or lat,lng.

You don't call tools by hand — the agent does. These are the kinds of questions that map onto a single call:

Discover the tools:

curl -sX POST https://data.sierragridteam.org/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

Ask what's happening near a town:

curl -sX POST https://data.sierragridteam.org/mcp \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc": "2.0", "id": 3, "method": "tools/call",
    "params": {
      "name": "grid_situation",
      "arguments": { "location": "Arnold, CA" }
    }
  }'

A trimmed response — note the preserved honesty fields:

{
  "resolved_place": "arnold",
  "mode": "WATCH",
  "summary": {
    "highest_severity": "SEVERE",
    "active_evacuations": null,          // null = UNKNOWN (Cal OES errored), not 0
    "evacuation_status": "UNAVAILABLE"
  },
  "domains": [ { "domain": "fire", "status": "OK", "active_count": 1 }, ... ],
  "disclaimer": "Reference only — verify with official sources ..."
}

Handshake (what an MCP client sends first):

curl -sX POST https://data.sierragridteam.org/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize"}'

This is reference-only, life-safety data. Every result carries a disclaimer, and an agent relaying it must respect the fail-loud contract:

• Every source reports OK / STALE / UNAVAILABLE. UNAVAILABLE means the status is unknown, not clear — never present absence of data as an all-clear.
• Evacuations: active_evacuations is an explicit null (source errored → unknown), 0 (Cal OES healthy, no active zones — caveated, not a guarantee), or N (active). null and 0 are different — never collapse them.
• Evacuation orders and directive text are shown verbatim; link the authoritative source, don't paraphrase.

The full model — severity scale, place ids, the evacuation contract — is in the /api/v1 reference. The MCP server also exposes it as an in-band resource (grid://reference).