EchoAct

EchoAct reads Korean and English text aloud on your own computer. Nothing you type is sent anywhere: the model runs on your CPU, and once it is downloaded the app works with no network connection at all.

• Read a document, follow along. The sentence being played is marked on screen. The mark is drawn as an outline around the letters rather than as a heavier weight, so the text never shifts under you while you read.
• Playback starts before generation finishes. The first sentence is spoken as soon as it is ready, and the rest follows.
• It stays out of the way. You choose how much processor and memory the app may use, and it stays inside that while you work on something else.
• Automations can use it too. A local HTTP service on the loopback address, and an MCP server for tools that speak that protocol. Both are the same engine, the same limits, and the same one job at a time.

Windows 11 and macOS 14 (Apple Silicon). No Python, Node, or database server to install.

Run it with one command

With uv installed, no checkout or venv needed:

uvx echoact

uv fetches EchoAct into an isolated environment and starts the app. The MCP server is the same package: uvx --from echoact echoact-mcp.

Running it from a checkout

uv sync
uv run python -m echoact

The first launch offers to download the speech model (about 385 MB) and shows its licence terms before it does. Generation works offline afterwards.

Local service

The service is on by default and listens only on 127.0.0.1:8765. It always requires a credential; one is created for you on first launch and is shown in the app under Settings. It cannot be reached from another machine, and there is no way to configure it to be.

Turn it off in Settings if you do not want it. The app, generation, playback and the library all work exactly the same with it off.

One generation runs at a time across the app and every client, so a second request is refused with a retry-after hint rather than queued.

MCP

Off by default; enable it in Settings. The MCP server is a small process your MCP client starts, which talks to the local service over loopback with a credential you issue for it. It adds nothing the HTTP service does not already do, and its permissions are exactly that credential's.

Connect an app in the header writes the configuration for you: a tab each for Claude Desktop, Claude Code, Codex CLI, VS Code and Cursor, with the right key name and file path for each, the command this particular installation actually uses, and a Copy button. Issue a credential from that screen and it goes straight into the snippet — the only moment it can, since the app keeps only a verifier.

Tools: list_models, estimate_speech, create_speech, get_speech_job, cancel_speech_job, list_speech_segments, get_speech_result, list_speech_history.

Your data

Documents and generation history are saved only when you ask. Results from a one-off job are cleaned up when they are replaced or when the app exits, and a result produced for an integration stays retrievable for an hour. Logs contain no text and no audio. Nothing is uploaded, and there is no telemetry.

Repository

docs/design.md    The requirements baseline. Every behaviour traces to an
                  identifier in it; Appendix A records what was measured.
echoact/          The application.
spikes/           Standalone measurements that settled open questions.
tests/            pytest. `-m engine` needs the model; the rest does not.

CLAUDE.md is the contributor's short version.

Author

Huijae Lee (ZeroAct) — lhj56822@gmail.com — github.com/ZeroAct

Licence

The application code is MIT (see LICENSE). The speech model is downloaded separately and carries its own licence with use restrictions, which the app presents for acceptance before it prepares the model for the first time. Being able to run it locally is not a right to redistribute it.