Skip to main content
Koharu Agent is a project assistant inside the editor. It runs on any model you can pick for translation that calls tools: a hosted model, a subscription such as ChatGPT, or one of Koharu’s general local models.
1

Connect a provider

Add an API key or sign in under Settings → Providers, or skip this step to use a local model. See translation providers.
2

Open the Agent panel

Open a project and select the Agent tab beside Properties on the right.
3

Choose a model

Select a model and a reasoning level at the top of the panel. Provider default uses the model’s own level. More reasoning can help with ambiguous edits but takes longer.

What context is shared

Each request starts with the open project’s outline: its pages and their layers. The Agent renders and attaches a page image only when it calls its page-view tool; it does not attach every page. A hosted model’s provider receives what the Agent sends it.
The Agent and translation share provider connections, but not models: the Agent’s model and reasoning are saved in the [agent] section of ~/.koharu/config.toml, apart from translation’s. That section also holds the Agent’s system prompt, max_turns (20 model calls per request by default), and corrections (how many times a malformed tool call goes back to the model, 2 by default).

Connect other agents over MCP

While Koharu runs, it serves the same tools to agents in other programs over MCP’s Streamable HTTP transport at http://127.0.0.1:7878/mcp, shown at the top of the Agent panel. Add that URL to an MCP client on the same machine. The server works on whichever project is open, accepts connections from this machine only, and refuses requests from web pages.

Work with a project

Give the Agent a precise scope, review its edits, and use project history.