How Zephyr works, end to end.

A request and a reply

editor_write takes a path and text. It writes to the buffer, not straight to disk, so unsaved changes and formatting stay intact. That is the one behavior worth knowing: nothing the CLI does bypasses what you see in the editor.

POST http://127.0.0.1:9222/mcp
Authorization: Bearer $ZEPHYR_TOKEN

{"jsonrpc":"2.0","id":7,"method":"editor_write",
 "params":{"path":"src/main.rs","text":"fn main() {}\n"}}

{"jsonrpc":"2.0","id":7,
 "result":{"ok":true,"bytes":13,"buffer":true}}

Health, and turning it off

/health is the one route that needs no token and does no UI work, so a CLI can check whether the editor is up before it tries anything else.

GET http://127.0.0.1:9222/health

{"status":"ok","version":"1.1.11","pid":18422}

A single switch in Settings stops the listener. The token is written next to the settings, and a startup call that arrives before the window is ready waits briefly and retries rather than failing outright.

The methods

Grouped by what they touch. The full list is longer than this; the editor, terminal, and git groups cover most sessions.

editor_read, editor_write, editor_insert, editor_replace
editor_search, editor_open, editor_active_file
terminal_create, terminal_send, terminal_read, terminal_close
git_status, git_diff, git_stage, git_commit, git_branch
git_log, git_blame, git_stash, git_conflicts
window_focus, panel_open, command_run, settings_read

One-click registration

Settings has a registration row for each supported CLI. It writes the config, keeps the token in sync, and tells you the exact file it changed.

• Claude Code
• Codex
• Gemini CLI
• opencode

{
  "mcpServers": {
    "zephyr": {
      "type": "http",
      "url": "http://127.0.0.1:9222/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Three adapter formats

Pick OpenAI, Anthropic, or Gemini and the request shape, the tool-call encoding, and the streaming parser all follow that provider. Inline XML tool calls from a local model work too: DSML, tool_calls, and antml invoke are all parsed, and the tags never reach the visible reply.

Local endpoints

Anything that speaks the OpenAI format works. Ollama is the common case: set the base URL and the model name, leave the key blank.

provider:  ollama
base_url: http://127.0.0.1:11434/v1
model:     qwen2.5-coder:14b
api_key:   (leave blank)

Parallel subagents run against whichever provider is configured, and an optional local RAG index can be built over the workspace without anything leaving the disk.

Run it

git clone https://github.com/ShinRyu04/Zephyr
cd Zephyr
npm install
npm run tauri dev

Dev mode hot-reloads the front end and restarts the Rust side when it changes.

Build a release

npm run tauri build

Installers land in the Tauri bundle output directory, one per target you enabled. The Windows target is the one that gets built and tested; Linux bundles are best-effort.