cookbook / local models

Tellie for local-model tools (LM Studio, OpenClaw)

Running models on your own Mac, LM Studio, OpenClaw, an Ollama-driven agent, can push to your notch exactly like Claude does. Pick the path that matches the tool: MCP for model runners, the CLI or URL scheme for agent frameworks. Needs Tellie. No accounts, no keys, all on-device.

LM Studio (and any MCP client)

LM Studio speaks MCP, so one command wires Tellie in as a server:

npx -y @tellie/cli setup lmstudio

That adds Tellie to ~/.lmstudio/mcp.json (it backs up first and leaves your other servers alone). Then: fully quit and reopen LM Studio so it re-reads the config, load a tool-calling model on the same Mac, and in the chat click the wrench icon in the message bar to switch mcp/tellie on for the conversation. Ask it to "call the update_status tool," approve the tool call, and the strip drops into your notch. Three tools appear: update_status, flash_status, send_to_tellie. Undo anytime with npx -y @tellie/cli setup lmstudio --off.

GUI apps need the full path to npx. LM Studio (and most desktop MCP clients) launch with a minimal PATH that doesn't include nvm or Homebrew node, so a bare "npx" won't spawn and the server silently never appears. The setup lmstudio command writes the absolute path for you. If you wire it by hand, run which npx and use that full path as the "command".

Prefer to wire it by hand, or using Claude Desktop / Cursor / another MCP client? Drop this into that client's MCP config:

{
  "mcpServers": {
    "tellie": { "command": "npx", "args": ["-y", "@tellie/mcp"] }
  }
}

OpenClaw (and any agent that runs commands)

An OpenClaw agent already has everything it needs: a shell. There's nothing to wire, just have it run the CLI or the raw URL scheme.

tellie update "Refactoring the auth module…" --source OpenClaw --icon hammer
tellie flash  "Done" --source OpenClaw --icon checkmark.circle

The hands-off way: point the agent at tellieapp.com/llms.txt and say "Tellie me." It learns the scheme cold (that is how the OpenClaw agent Storm picked it up and made it a habit). For a permanent habit, add one Tellie line to the agent's standing instructions, so it narrates without being asked.

The model has to be able to call tools

For the MCP path, the local model needs to be reasonably good at tool-calling. Qwen, Llama 3.x instruct, and similar handle it well; very small or base models may not reliably make the call. If tool-calling is shaky, drive Tellie from the harness instead: run the tellie CLI around the model call. That path does not depend on the model at all, which is the whole point of a generic primitive.

Good to know

• It's all free. The MCP server, the CLI, and the tellie:// URL scheme are the free developer surface. Persistence (Pulse Log) and pro surfaces are the paid part.
• The URL scheme is the universal fallback. Anything that can run open -g "tellie://update?text=…" can drive the notch, no install, no MCP, no CLI.
• Same config, other clients. The @tellie/mcp block works in Claude Desktop, Cursor, and other MCP clients too; only the config file path differs.