Generate narration from Terminal, or let your coding agent create, inspect and organize audio. It all runs on your Mac, with the models and voices already there.
Set up automationFree to download, Pro free for 3 days · 7-day refund · macOS 15+ on Apple Silicon
murmur generate --input video/script.md --output video/voiceover.wav
Ready for Codex, Claude Code, Cursor, and scripts
Introducing automation
Murmur is not limited to the app window. The same local engine that powers the studio answers to Terminal and to the coding agent you already use.
Every automation rides the app's own pipeline. Your installed models, saved voices and license stay in one place, and generation never leaves your Mac.
Two surfaces
Scripts call the murmur command. Agents call the MCP server. Both hand the work to the Mac app, which owns the models.
Turn this scene into a WAV I can drop in the edit
murmur generate --text "Scene one." --voice af_bella --output audio/01.wav
Example output: audio/01.wav
murmur batch --manifest video/scenes.json
The murmur CLI. Turn any text or Markdown script into WAV or M4A with one command, batch a whole project from a small JSON manifest you can version in git, and install, select, inspect and remove local models from scripts. Command reference
Check Murmur is ready, then voice the three scenes
murmur_statusmurmur_list_voicesmurmur_generate_batchmurmur_get_jobCodexClaude CodeCursorAny MCP client
The MCP server. Codex, Claude Code and Cursor manage models, discover voices and create audio as tool calls. It runs over stdio from the signed helper inside the app, so there is no server to run. Connect an agent
Workflow
Turn on local automation in Settings, Automation and click Install CLI. Murmur links the signed helper to ~/.local/bin/murmur.
Pass text directly or point at a script file. Pick a voice and speed, choose WAV or M4A, and Murmur renders on-device.
Add one MCP server entry to Codex, Claude Code, or Cursor. The agent gains ten Murmur tools for models, voices, generation, and jobs.
Keep a scenes.json beside your video script. Relative paths resolve from the manifest folder, so the whole voiceover versions with the project.
# 01 after Install CLI in Settings
murmur status
# 02 generate from a script
murmur generate --input scripts/chapter-01.txt --model kokoro --output audio/chapter-01.m4a
# 03 connect Claude Code
claude mcp add --transport stdio --scope user murmur -- /Applications/Murmur.app/Contents/Helpers/murmur mcp serve
# 04 batch a whole voiceover
murmur batch --manifest video/scenes.json
Agent workflow
With MCP connected, your agent can install and select a model, inspect its voices, queue a batch, poll the job, and hand back the exact files it created. Your script never leaves the workspace.
See the agent workflowRead video/script.md. Split it into intro, demo, and outro. Use Murmur to create WAV files in video/voiceover/, keeping one voice and pace across every scene. Don't replace existing takes.
video/voiceover/01-intro.wavvideo/voiceover/02-demo.wavvideo/voiceover/03-outro.wavExample prompt and output.
MCP tools
Manage local models, discover voices, queue audio, and track or cancel jobs without opening a cloud dashboard.
murmur_statusSetup, license, and automation readiness.murmur_list_modelsEvery model, capability, and install state.murmur_list_voicesPreset voices and saved reference profiles.murmur_install_modelDownload a model as a trackable job.murmur_set_default_modelChoose the installed default model.murmur_remove_modelRemove downloaded model files with confirmation.murmur_generate_speechQueue one WAV or M4A inside the workspace.murmur_generate_batchQueue a sequence of named files as one job.murmur_get_jobProgress and finished artifact paths.murmur_cancel_jobStop any active long-running job.Local by design
The same privacy story as the app, enforced at the helper level, so a script or an agent can only do what you would do yourself.
Read the safety notesAutomation starts disabled. Nothing can call the helper until you allow it inside Murmur.
The Mac app runs every generation, so the CLI and agents never start competing model servers.
MCP reads and writes only inside the directory where the agent launched it.
Existing output fails safely unless you pass --force on purpose.
Removing downloaded model files requires confirmation, and bundled Kokoro cannot be removed.
Prompts, scripts, saved voices, and generated audio remain on your Mac.
Install the CLI from Settings, then let Terminal, Codex, Claude Code and Cursor narrate with the voices on your Mac. Open the automation docs
Free to download, Pro free for 3 days · 7-day refund · macOS 15+ on Apple Silicon
