Agent Speak

Give VS Code agents a local, policy-controlled voice. Agent Speak registers a bundled MCP server that can speak text, play approved audio cues, report playback status, and cancel playback through your desktop audio output.

No separate Agent Speak installation or mcp.json entry is required. The extension creates an isolated profile automatically and keeps its runtime, configuration, providers, and audio on the VS Code desktop host.

Get started

1. Install Agent Speak and reload VS Code if prompted.
2. Open an agent-capable chat in VS Code and ask it to use Agent Speak. The agent can inspect get_audio_capabilities before calling speak_text.
3. Open the extension's gear menu and choose Settings to adjust permissions, voice, output, playback policy, or audio cues.

The default profile enables local text-to-speech with a conservative gain and queued playback. macOS and Windows use their built-in system TTS. The Linux x64 package includes a self-contained eSpeak NG provider, so first speech does not require a separate voice or model download.

Tool use and approval remain controlled by the VS Code MCP host. Hearing speech also provides a useful end-to-end check that the server, policy, and local audio output are working.

Human-friendly settings

The Settings panel provides structured controls for:

• arbitrary text and local-audio permissions;
• system voice and approved output selection;
• gain, duration, queue, interrupt, and bounded mixing policies;
• TTS enablement and text limits;
• speech and local-file audio cues, including add, remove, and reorder;
• recent playback states, targeted cancellation, and emergency stop.

Changes remain drafts until Apply Changes validates and atomically installs the complete profile. Stale or conflicting edits do not replace the last valid configuration. Open config.toml remains available for advanced provider, output, and history settings, and the extension validates that file when it is saved in VS Code.

Local and Remote SSH

Agent Speak runs as a local UI extension. In a Remote SSH window, the workspace and terminal can be remote while the bundled runtime, managed profile, provider processes, MCP server, and audio output stay on your desktop. Nothing needs to be installed on the SSH host.

Use Agent Speak: Show Host Diagnostics to inspect the effective desktop-host paths and runtime target. The managed extension profile is separate from any standalone Agent Speak installation on the same computer.

Privacy and control

• The default TTS path is local. Text leaves the machine only if you explicitly configure a network-backed provider.
• Spoken-text history is disabled by default.
• Arbitrary local-file playback is disabled by default. Enabling it permits the agent to request accessible absolute local files, subject to the profile's playback limits.
• The Settings playback view receives lifecycle IDs and states, never spoken text, audio paths, provider options, or the private control token.
• Emergency stop discards queued items and stops active playback for the extension-owned MCP session.

Supported packages

• macOS ARM64
• macOS x64
• Windows x64
• Linux x64

VS Code 1.101 or newer is required. VS Code Desktop local and Remote SSH windows are supported. WSL, Dev Containers, and Codespaces Desktop have not yet completed local-playback acceptance.

Providers and advanced configuration

The bundled default voice is designed to work immediately. Optional utterpipe-* providers can be placed in the folder opened by Agent Speak: Open Provider Folder, then selected through config.toml. Provider-specific models, voices, options, network behavior, and licenses are governed by that provider.

See the repository for development documentation, source, checksums, release artifacts, and issue tracking. Agent Speak core configuration is documented in the Agent Speak repository.

Release integrity

The bundled native runtimes are currently unsigned and are not notarized. Target-specific VSIX files and SHA-256 checksums are published with each GitHub release, and the extension carries the required third-party notices.

Report problems through GitHub Issues.