mireye-mcp

Developed in a private monorepo. This repository (Mireye-Labs/mireye-earth-mcp) is the published source of the mireye-mcp PyPI package, snapshot-synced on every release. Issues and PRs are welcome here (changes are applied upstream and re-synced; the test suite runs in the monorepo); releases land via PyPI.

Expose Mireye Earth's /v1/ask and /v1/fetch endpoints to MCP clients that need a local stdio adapter (Claude Desktop, Cursor, custom agents built on the mcp Python SDK) as native tools — no HTTP wiring required.

This is a standalone PyPI package (mireye-mcp) with only two runtime dependencies — httpx and mcp. It does not pull in the geospatial backend, so there's no GDAL / rasterio / DuckDB build step.

The server is a local stdio adapter with no geospatial business logic: tool handlers POST to the deployed Mireye HTTP API, while read-only MCP resources fetch and cache the public field catalog.

Mireye also exposes a hosted remote MCP endpoint at https://api.mireye.com/mcp for clients that support Streamable HTTP and native OAuth. Use that hosted endpoint for Claude Code so /mcp opens the browser sign-in flow. This package remains the local stdio path; it uses mireye-mcp login or MIREYE_BEARER_TOKEN for credentials.

---

What the agent gets

Two tools, both prefixed mireye_ so they sort together and don't collide with generic ask / fetch tools from other MCP servers:

Catalog context is exposed as MCP resources instead of extra tools:

Workflow prompts are also registered: mireye_ask, mireye_fetch, mireye_site_report, mireye_flood_check, mireye_wildfire_underwrite, and mireye_pick_fields. Claude Code surfaces MCP prompts as slash commands under the form /mcp__<server>__<prompt>.

There is no third list_fields tool. Agents that need the catalog should read the MCP resources above; the stdio adapter backs them with GET /v1/meta/fields and a 1-hour ETag-aware cache.

---

Install local stdio — one command with uvx

uvx mireye-mcp

That's it. uvx (bundled with uv) downloads the package into a managed cache, runs the entry point, and the next invocation is instant. No venv to manage, no pip install, no native builds.

If you don't have uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

Plain pip also works:

pip install mireye-mcp
mireye-mcp  # entry point

---

Official MCP Registry

This server publishes to the Official MCP Registry as com.mireye/earth (the entry goes live with the next tagged release; the publish job ships from release.yml). The registry entry carries both distributions:

• the PyPI package mireye-mcp (local stdio, run via uvx), and
• the hosted remote https://api.mireye.com/mcp (Streamable HTTP + OAuth) for clients that prefer a remote server.

Once the entry is live, registry-aware clients (VS Code, the GitHub MCP Registry, and anything else that reads the official registry) can discover and install it from there — no manual config required.

---

Wire it into Claude Desktop

First authenticate the local adapter:

mireye-mcp login

For non-interactive hosts, set MIREYE_BEARER_TOKEN instead.

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "mireye-earth": {
      "command": "uvx",
      "args": ["mireye-mcp"]
    }
  }
}

Restart Claude Desktop. The two tools (mireye_ask, mireye_fetch) appear under the 🔌 menu, with catalog resources and prompts available to clients that surface those MCP primitives.

To point at a self-hosted deployment instead of the default Fly URL:

{
  "mcpServers": {
    "mireye-earth": {
      "command": "uvx",
      "args": ["mireye-mcp"],
      "env": {
        "MIREYE_BASE_URL": "https://your-deploy.example.com"
      }
    }
  }
}

---

Wire it into Claude Code

Use the hosted HTTP MCP endpoint instead of this local stdio package:

claude mcp remove mireye-earth -s user   # only needed if an old stdio entry exists
claude mcp add --transport http --scope user mireye-earth https://api.mireye.com/mcp

Restart Claude Code, run /mcp, and follow the browser OAuth flow. Slash commands appear as:

• /mcp__mireye-earth__mireye_ask <lat> <lng> <question>
• /mcp__mireye-earth__mireye_fetch <lat> <lng> [fields] [preset]

Or just chat naturally — Claude Code will call the tool when relevant.

---

Wire it into Cursor

Cursor's MCP config lives at ~/.cursor/mcp.json (global) or <repo>/.cursor/mcp.json (workspace). Same shape as Claude Desktop:

{
  "mcpServers": {
    "mireye-earth": {
      "command": "uvx",
      "args": ["mireye-mcp"]
    }
  }
}

Open Cursor → Settings → MCP and confirm mireye-earth shows the two green tools. Catalog resources and prompts appear when the client supports those MCP primitives.

---

Wire it into a custom agent

If you're building an MCP client with the mcp Python SDK, point its StdioServerParameters at the installed entry point:

from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

params = StdioServerParameters(command="uvx", args=["mireye-mcp"])

async with stdio_client(params) as (read, write):
    async with ClientSession(read, write) as session:
        await session.initialize()
        tools = await session.list_tools()
        # tools.tools is [mireye_ask, mireye_fetch]
        resources = await session.list_resources()
        prompts = await session.list_prompts()

        result = await session.call_tool(
            "mireye_ask",
            {"lat": 40.7128, "lng": -74.0060, "question": "elevation?"},
        )

---

Configuration

Authentication

The local stdio adapter does not perform native MCP OAuth discovery. The HTTP API requires bearer auth for /v1/ask and /v1/fetch, so run the device login helper once:

mireye-mcp login

The command prints a verification URL and code, waits for approval in the Mireye account page, and stores a Mireye API token locally. You can also provide a token directly:

{
  "mcpServers": {
    "mireye-earth": {
      "command": "uvx",
      "args": ["mireye-mcp"],
      "env": {
        "MIREYE_BEARER_TOKEN": "eyJ..."
      }
    }
  }
}

Check or remove local credentials:

mireye-mcp status
mireye-mcp logout
mireye-mcp logout --revoke

Credentials are bound to their base URL

login records the MIREYE_BASE_URL it ran against, and the stored token is only ever sent to that same URL. If MIREYE_BASE_URL later points somewhere else, tool calls behave as logged out and the error names both URLs — re-run mireye-mcp login against the new URL, or set MIREYE_BEARER_TOKEN explicitly. Credentials files without a recorded base_url (e.g. hand-written) are treated as bound to the default https://api.mireye.com.

Two more guardrails:

• Tokens of any kind are never sent over plain http://, except to loopback hosts (localhost / 127.0.0.1 / [::1]) for local development.
• status and logout --revoke operate on the stored base URL, so you can always inspect or revoke a stored login even while MIREYE_BASE_URL points elsewhere.

For native MCP OAuth, configure your client to use the hosted remote MCP URL https://api.mireye.com/mcp instead of launching this stdio binary. The remote endpoint advertises OAuth metadata and uses browser OAuth 2.1 + PKCE.

---

Troubleshooting

Server doesn't appear in Claude Desktop. Check that uvx resolves on the PATH used by the GUI app (macOS launches GUI apps with a minimal PATH). Test from a terminal: which uvx. If empty, install uv: curl -LsSf https://astral.sh/uv/install.sh | sh. If uvx lives at /Users/you/.local/bin/uvx, use the absolute path in command.

Tools not appearing under the 🔌 menu after restart. Watch ~/Library/Logs/Claude/mcp-server-mireye-earth.log. The server logs to stderr on startup; you should see [mireye-mcp] starting base_url=…. If there's no log, uvx isn't being invoked — usually a PATH issue.

ConnectError / ReadTimeout on the first tool call. The hosted API keeps its machines running, but calls right after a backend deploy can be slow while geospatial sources warm in the background, and some fields depend on slow upstream federal services. Increase MIREYE_TIMEOUT_S to 90 if you are repeatedly hitting timeouts.

HTTP 400 coord_out_of_bounds. Mireye is US-only in V1. The accepted envelope is lat ∈ [18, 72], lng ∈ [-180, -65] — covering the lower 48, Alaska, Hawaii, and US territories.

HTTP 400 fields_unknown. The field name is not in the catalog. Hit https://api.mireye.com/v1/meta/fields to see the canonical list. Common surprises: elevation_m is elevation, floodplain is within_floodplain_polygon.

HTTP 4xx/5xx in general. Tool errors include actionable JSON fields such as code, message, http_status, request_id, tool, and retryable. Agents should retry only when retryable is true.

Auth errors. 401 means the MCP server has no token, the token expired, or the token was revoked. Run mireye-mcp login again, or set MIREYE_BEARER_TOKEN. 403 means the signed-in account is not allowed by the backend account policy.

---

What this server does NOT do

• No data-result caching. The HTTP API has its own response cache (local disk, plus a shared Redis tier in production). The stdio adapter only caches the public field catalog resource.
• No geospatial business logic. The adapter validates obvious MCP input bounds and field-count limits, but source orchestration stays in the API.
• No streaming. MCP streaming responses are V1.5.
• No Firebase validation in MCP. The MCP package forwards bearer tokens; the HTTP API owns token verification and account policy.
• No in-process imports of mireye_earth. The server talks to the API over HTTP only, so it ships as a separate slim PyPI package (mireye-mcp) and can be installed without the data backend.

---

Releasing

This package is published from this directory on every v* tag on the main repo. To release locally:

cd mcp_server
python -m build
twine upload dist/*