Ultimate Local Agent – Private AI Coding Agent on Local Models

# Local Qwen Ultimate Agent

A private, Windows-native project for running Qwen Code through a local Ollama model without paid API credits, then adding isolated browser and computer-control capabilities in verified stages.

## Features

- Qwen Code with local Ollama inference

- Qwen 3.5 2B Q4_K_M model with an 8K working context

- Local adapter that disables expensive hidden reasoning while preserving tools

- Default approval mode for file edits, shell commands, and browser actions

- Visible Playwright browser in a fresh in-memory Edge profile

- No modification of global Qwen or OpenCode provider settings

Start the agent with `outputs/qwen-local.cmd` from the project directory where it should work.

## Browser control

The launcher supplies pinned Playwright MCP 0.0.83 explicitly to Qwen. It opens a separate visible Edge window, discards cookies and local storage when the session closes, and leaves Qwen in default approval mode. Every MCP browser action currently asks for approval.

See:

## ICM personal coding controller

Clone with submodules, or initialize them in an existing clone:

git clone --recurse-submodules

git submodule update --init --recursive

From the project root in PowerShell 7, use the `.ps1` launcher. Each command writes one JSON object to stdout; errors set a nonzero exit code and return `ok: false`. The `.cmd` launcher supports the same commands in Command Prompt; there, escape JSON quotes as `\"` (for example, `outputs\icm-agent.cmd route --profile "{\"deterministic\":true}"`).

outputs\icm-agent.ps1 validate

outputs\icm-agent.ps1 route --profile '{"deterministic":true}'

outputs\icm-agent.ps1 run create --id research-example --objective "Find provider information" --profile '{"externalResearch":true,"writesCode":false}'

outputs\icm-agent.ps1 run status --id research-example

outputs\icm-agent.ps1 sync preflight

outputs\icm-agent.ps1 sync verify

The controller creates an editable run under `runs/<id>/`. For an Instant route, use `research/instant-request.md` as a visible inbox and record the answer, source URL, and retrieval date in `research/instant-response.md`. The free ChatGPT UI is a manual handoff, not an automated API.

`sync preflight` without a run ID fetches and rejects dirty trees, missing or unexpected tracking, and unequal commits. `sync verify` without a run ID fetches and checks tracking and commit equality; it does not reject dirty trees. These two standalone commands are diagnostics and have no run in which to record recovery.

For a workflow transaction, first create the run, then acquire its lock:

(Additional install and ICM controller docs live in the project.)