Talk to your agents

Point an agent at Earshot and it can ask you things out loud.

Agents reach Earshot through an MCP server inside the app bundle, over a local socket. Nothing goes over a network and there is nothing to sign in to. Setup registers it for you.

  1. 1

    Open Setup from the menu bar

    It looks for the agents you already have installed and lists them. Claude Code, Claude Desktop, Codex, Cursor, opencode, Windsurf, Zed, Cline, Gemini CLI and Hermes are recognised by name.

  2. 2

    Press Register next to the ones you want

    That writes an MCP server entry into the agent's own config, pointing at the server inside the Earshot bundle. Nothing is installed and nothing runs until an agent calls it.

  3. 3

    Restart the agent

    Every one of them reads its config at startup. This is the step people miss, and the symptom is an agent that simply never mentions Earshot.

  4. 4

    Add the speak-mode instructions, if you want it talking by default

    Setup offers to write a short block into the agent's guide file, telling it to end each turn out loud and to ask rather than assume. Without it the tools are available and most agents will not reach for them.

The three tools

Three, and no more than three. A larger surface would mean an agent choosing between eight things it half understands. This is what each one puts in the notch.

ask_user(question, choices?)

Speaks the question, opens the microphone, and blocks until you answer. Your words come back as the tool result, so the agent never has to end its turn to hear you. The panel turns cool the moment the floor is yours.

choices is optional and never read out. The agent phrases its own options into the question, and the list is only what your answer is matched against. Anything matching none of them comes back word for word, because an agent handed the wrong option confidently has no way of noticing.

speak(text)

Says a short summary out loud. Meant for the end of a turn, not for reading code: anything longer than a couple of sentences is faster to read than to hear.

working(status)

Puts a phrase in the notch without saying anything. It is silent, it costs nothing, and it is the whole difference between a long pause and a session that looks dead.

Cutting it off

Hold the hotkey while it is talking and it stops, immediately, and the floor is yours. You never have to wait out a summary you have already understood.

Interrupting a spoken replyHolding the key stops it mid-sentence and opens the microphone.

When it goes wrong

The agent does not know Earshot exists

It has not been restarted since the server was registered.

Quit it properly and open it again. If it still cannot see it, Setup shows whether the config it wrote is the one that agent actually reads.

My agent is not in the list

It is not one of the ones detected by name, which is most of them: there are a great many and the protocol is the same for all.

Setup scans for config files it recognises the shape of and offers to add the server to those too. Failing that, the Anything else tab gives you the block to paste by hand.

It asks a question and then answers it itself

The agent called speak rather than ask_user, so nothing ever opened the microphone.

Add the speak-mode block from Setup. It is the part that tells an agent a question means ask_user.