Ollama launch: integration handoff

A new launcher flow that hands local models into tools like Codex and OpenCode.

Tags
Layout

Key takeaways

  • ollama launch turns model selection into a single handoff step.
  • The integration list is the contract: install the tool before you launch it.
  • A short, visible flow beats hidden config when you are new to local stacks.

This experiment is about a new entry point: ollama launch. Instead of wiring a tool to a model by hand, the launcher asks which integration you want and then routes the model into it. The result feels like a simple handoff layer for local AI tools.

Architecture map

ollama launch acts like a dispatcher. You pick a tool, then a model, and it hands the model into that tool.

Ollama launch handoff Ollama launch selects an integration, then hands a model into the tool session. Ollama Launch menu Tool A Tool B
A small handoff layer: choose the tool, then the model.

What happened

The flow started with a simple command and a menu. The launcher listed available integrations (Claude, Codex, OpenCode, and others), then asked for a model. When I selected Codex without having it installed, the launcher surfaced the missing dependency clearly and told me what to install.

Ollama launch integration menu listing Claude, Codex, Droid, and OpenCode.
Integration menu: the launcher presents the available tools.
More screenshots
Ollama launch model selection screen showing multiple local models.
Model selection: choose a local model for the selected tool.

The two gates

The first gate is tool installation. If the integration is not installed, the launcher stops and tells you exactly what is missing.

The second gate is model availability. If the model is not local, the handoff stalls until it is pulled or selected.

Setup walkthrough

  1. Run the launcher.
  2. Select an integration (Codex, OpenCode, or another tool you have installed).
  3. Select a local model.
  4. Start the tool session.
ollama launch

First-time config

When Codex was missing, the launcher pointed to the installation command. After installing, the same launch flow worked cleanly.

npm install -g @openai/codex
Terminal message showing Codex is not installed and suggesting npm install -g @openai/codex.
A clean failure path: missing integrations are explicit.

Quick checks

ollama launch codex
ollama launch opencode
ollama launch claude
Terminal showing Codex running after ollama launch codex.
The handoff works once the tool is installed.

Failure modes

  • Launching a tool before it is installed.
  • Selecting a model that is not available locally.
  • Assuming the launcher is a model manager rather than a handoff layer.

What made the difference

The launcher compresses setup into one visible decision tree. Instead of hunting for flags and configs, you pick a tool and a model and move on.

What I would do next time

I would pre-install the integrations I care about, then treat ollama launch as the default entry point for local tool sessions.

Continue Previous Soothsayer MCP kernel: from prompts to controlled orchestration Next AWS GPU instance baseline
Back to /shelf/local-experiments Related