Client Setup¶
Paste the config block for your agent. You only need to do this once.
The baseline (uvx timesfm-mcp) works identically in every client below — no extra install needed.
Claude Desktop¶
Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
Restart Claude Desktop. You'll see "forecast" appear in the MCP tools panel.
With TimesFM enabled¶
System requirement: ≥ 16 GB RAM · ~800 MB disk
TimesFM downloads ~800 MB of model weights on first use and requires at least
16 GB of RAM. Most users should use the uvx baseline config above —
it works on any machine with no download.
Then update your Claude Desktop config to use the installed binary instead of uvx:
Allow 1–2 minutes on the first invocation while model weights download and load.
Claude Code¶
Add to your project's .mcp.json, or run the CLI command:
Or manually in .mcp.json:
Verify it's active:
You should see forecast listed with tools forecast, list_backends, backtest.
GitHub Copilot (VS Code)¶
Requires VS Code 1.96+ with the GitHub Copilot extension and agent mode enabled.
Create .vscode/mcp.json in your workspace (or add to your VS Code user settings.json under the "mcp" key):
Open Copilot Chat, switch to Agent mode using the mode picker, and the forecast, list_backends, and backtest tools will be available.
Cursor¶
Open Settings → MCP (or ~/.cursor/mcp.json) and add:
Restart Cursor. The forecast tools are now available to Cursor's agent.
Windsurf (Codeium)¶
Edit ~/.codeium/windsurf/mcp_config.json:
Restart Windsurf. The tools will appear in the Cascade agent panel.
Cline (VS Code extension)¶
Open the Cline sidebar → MCP Servers tab → Edit MCP Settings. Add to cline_mcp_settings.json:
{
"mcpServers": {
"forecast": {
"command": "uvx",
"args": ["timesfm-mcp"],
"disabled": false,
"autoApprove": []
}
}
}
The forecast tools will appear in Cline's tool list immediately — no restart needed.
Continue.dev¶
Edit ~/.continue/config.json and add to the mcpServers array:
Reload the Continue extension. Use @forecast in the chat to invoke the tools.
Zed¶
Edit ~/.config/zed/settings.json and add under context_servers:
{
"context_servers": {
"forecast": {
"command": {
"path": "uvx",
"args": ["timesfm-mcp"]
},
"settings": {}
}
}
}
The forecast tools will be available in Zed's AI assistant panel.
Any other MCP client¶
The server speaks the standard MCP stdio transport. Start it with:
Set PORT to change the HTTP port:
Environment variables¶
| Variable | Default | Effect |
|---|---|---|
TIMESFM_MCP_BACKEND |
auto |
Set to baseline to force the statistical backend even when TimesFM is installed |
PORT |
8000 |
HTTP port (only used with --http) |
Troubleshooting¶
"No tools found" — Check that uvx is on your PATH (which uvx). Install with pip install uv or brew install uv.
TimesFM import error — Run pip install "timesfm-mcp[timesfm]" and restart the server. If pip install timesfm was previously run separately, uninstall it first (pip uninstall timesfm) to avoid the old 1.x version shadowing the bundled 2.5 source.
Server not responding — Check the MCP logs in your client. The server logs startup info to stderr.