Skip to content

kaimahi-agents/kmx-agent-builder

v0.9.0-rc.3

Create, edit and test agents in any repository with kmx in WSL and the Copilot canvas.

GitHub

kmx in Windows Copilot through WSL

The plugin contains skills, canvas assets and MCP configuration. It uses the Linux kmx binary in your default WSL distribution; it does not bundle an engine. The v0.9.0-rc.2 install below exposes Create/Edit/Run/Test. The matching current-checkout engine adds the revision-checked set_tool_state folder-settings writer; installed use needs a newer release containing that tool, not rc.2. Internal preview children retain only Create/Edit/Run/Test and never gain config-write authority.

Install

  1. In WSL, install the v0.9.0-rc.2 Linux kmx binary at /usr/local/bin/kmx:

    curl -fsSL https://raw.githubusercontent.com/kaimahi-agents/kaimahi/v0.9.0-rc.2/install.sh | sudo sh -s -- --version=v0.9.0-rc.2 --bin-dir=/usr/local/bin

    The installer downloads the binary and checks its published SHA256 over HTTPS; it does not verify the separate signature bundle.

  2. In a Windows terminal with Copilot installed:

    copilot plugin install kaimahi-agents/kaimahi:plugins/copilot-wsl

There is no folder or model to configure during installation. The fixed command is wsl.exe -e /usr/local/bin/kmx mcp. The WSL and native-bundle plugins share the name kmx-agent-builder; installing one replaces the other. The Copilot CLI warns that direct plugin installs are deprecated; this install still works, and a marketplace install will come later.

Use

In Copilot, ask it to create an agent through kmx and open the agent canvas. Give it a full WSL path when the file is outside the current directory. The same server handles files in unrelated repositories.

Choose a runtime model in the canvas: picker override → authored Model.Preferred → known chat model → gpt-5.6-terra. Model may be omitted; a picker choice never rewrites the agent file. The open canvas keeps results for two models: picking a model with a result for the same file revision restores it without rerunning. Older results stay grey, with their model and revision context. A newer single-test attempt takes precedence over an older full-run summary. A following default reads the chat model when Run/Test starts and freezes it through save/execution; the displayed model can lag until idle refresh. Choices load when the menu opens and are cached; catalog failure does not block editing or offline tests. Familiar IDs get display names and family groups; unknown IDs remain literal and execution always uses the exact ID. Selecting a model pins it, including the current default; use the following action to return to the agent file or automatic default. Older hosts without the optional current-model getter use the established default. Transient getter failures or metadata deadlines retain the last valid model; explicit unknown/auto responses clear it. Unbound real-model runs use Copilot by default; effective fake is deterministic and offline. kmx model list --provider copilot lists the available model names.

Edit prompts and tests in the canvas. Enter saves and tests; Shift+Enter inserts one line break. Long prompts scroll inside their box, with the save/run footer outside the scroll area. Saved agent files remain ordinary Markdown. Tests send prompts and inputs to the selected model. On hosts that forward native tool progress, test rows update as cases finish. The final full Test result remains authoritative. Local model comparison is lost when the canvas is closed and reopened. An unconfirmed execution requires reload rather than silently retrying. If the host has no canvas renderer, the same kmx tools remain available in text.

See the agent-file guide for authoring and checks, and the macOS/PATH plugin guide for the Mac install and legacy native bundles.