tometo

tometo

tometo runs the poteto loop locally. Code, judgment, panel, and swarm can all be local models: Ollama, llama.cpp, MLX, or an OpenAI-compatible server at 127.0.0.1.

Name the data. Reproduce before you fix. Prove the change on the surface a person uses. Put a stronger model on the judgment, and a fast one on the edit. The parent fans out. Children stop when their task is done.

tometo spawn stays on this machine until you pass --allow-cloud.

pstack carries that method on Cursor and Grok Bot. See NOTICE.md.

A full local stack

~/.tometo/config.json can name a model on every role. This example uses Qwen3.8, the current open 27B weights. qwen3.8:27b is the default 18GB Ollama tag. qwen3.8:27b-q8_0 is the 30GB quant. unsloth/Qwen3.8-27B-GGUF is the llama.cpp repo.

setup writes Ollama into code, judgment, and swarm when no grok, claude, opencode, or codex binary is on PATH. When one of those CLIs is installed, setup leaves code and judgment on it and adds each model from ollama list to the panel. Paste the file below to pin every role to a local model.

{
  "version": 1,
  "lane": "local",
  "host": "local",
  "approve": "edits",
  "roles": {
    "code": {"runner": "ollama", "model": "qwen3.8:27b", "effort": "high"},
    "judgment": {"runner": "ollama", "model": "qwen3.8:27b-q8_0", "effort": "high"},
    "panel": [
      {"runner": "ollama", "model": "qwen3.8:27b", "effort": "high"},
      {"runner": "llama-cli", "model": "unsloth/Qwen3.8-27B-GGUF", "effort": "high"}
    ],
    "swarm": {"runner": "ollama", "model": "qwen3.8:27b", "effort": "high"}
  }
}

ollama on code and swarm gets the tool loop. Judgment, panel, llama.cpp, MLX, and openai are one reply. bin/tometo doctor checks that each runner’s binary is on PATH. bin/tometo spawn --dry-run prints the argv and starts nothing.

An mlx seat uses the same shape, {"runner": "mlx", "model": "<repo>"}, with a repo mlx_lm.generate can load. The Apple Silicon upload mlx-community/Qwen3.8-27B-4bit is an mlx-vlm build.

Install

From this checkout:

bin/tometo link
bin/tometo setup --yes
bin/tometo doctor

link symlinks the skills into the skill folders of the hosts installed on this machine (~/.grok, ~/.claude, ~/.cursor, ~/.codex) and into ~/.agents/skills.

Grok can also load the checkout as a plugin:

grok plugin install /absolute/path/to/tometo --trust

Then start a session and run /setup-tometo, then /tometo.

Use

/tometo reads ~/.tometo/config.json, then hosts/<host>.md, then one playbook:

playbook	for
investigation	how something works, or whether a claim is true
bug-fix	reproduce, cause, smallest fix, run it again
feature	a new behavior, driven the way a user drives it
opening-a-pr	a proven diff, small commits, a PR body with what you ran

The contract those files share is docs/runtime.md.

A role whose runner is another product starts with:

bin/tometo spawn --role code --prompt-file /tmp/task.md --cwd "$PWD" --worktree

--dry-run prints the argv and does not start a model. setup leaves model unset, so each CLI uses the default you already signed in with. Pin one by editing the JSON.

Runners

setup records a CLI it can see on PATH: grok, claude, opencode, codex. It also puts each Ollama model on the panel. A llama-cli, mlx, or openai role is used once you name its model in the config. A full local assignment is in A full local stack. The fields are in docs/runtime.md. spawn --dry-run prints the argv. Pi has a host file. This runner does not call Pi’s subagent tool yet.

Tests

python3 tests/test_cli.py