In-editor chat

Providers and keys

Open the toolkit's Settings and find Chat model. Pick a provider, paste a key if it needs one, and choose a model. These settings only affect the built-in chat; external clients use their own models.

Keys are encrypted with Windows DPAPI for your user account on this machine. They are never written to disk in plain text, and Settings only shows the last four characters. To change an endpoint, use Custom endpoint under Advanced connection.

After you save a key, the model list fills from the provider. You can also type an exact model id. The chat needs a model that supports tool calls; Settings flags models that are known not to.

Local models

Choose Ollama or LM Studio to keep the conversation on your machine. No key is needed. Start the local server first, with a model loaded that supports tool calls. The Ollama default model is qwen2.5-coder:7b-instruct; pull it, or pick another from the list.

ollama pull qwen2.5-coder:7b-instruct

Small local models follow long tool chains less reliably than large hosted ones. Start with short, specific requests.

Inspect, Draft and Production

The chat composer has three modes. The mode sets what the next run may do.

• Inspect reads and explains. It has read-only tools and makes no project changes. New chats start here.
• Draft can change the project. Each change waits for your approval in the chat before it runs.
• Production also sends every write through the plugin's exact review, even when Review each external edit is off. On this version that review covers only edits to one existing actor, so Production refuses other writes. See What can be approved.

The chat's server reaches the editor through the same plugin router as an external client. On this version the router applies Review each external edit to every write it receives, including the chat's. Turn it off for creation work in Draft.

How the chat runs

When the editor starts, the plugin launches a copy of the MCP server for the chat. It listens on http://127.0.0.1:7821. If something is already listening there, the plugin reuses it rather than starting a second one.

The plugin finds this server and Node.js like this:

• The server: SidecarEntryPath if you set it; otherwise the built server next to a linked plugin, mcp-tools/hayba-mcp/dist/index.js.
• Node.js: C:\Program Files\nodejs\node.exe or C:\Program Files (x86)\nodejs\node.exe. The plugin does not search your PATH.

If the chat cannot send, check the Output Log for LogHaybaMCP lines such as Node.js not found. or MCP server not found at:. See Node.js not found.

The chat server only accepts requests addressed to this machine; it refuses other Host and Origin headers. Chat sessions are stored with secrets redacted.