Claude Code
Sends follow-ups into the sessions you already have open and starts new ones headless. Risky tools stay blocked until you say so.
Open source · macOS · bring your own keys
OpenMimir sits in front of Claude Code, Codex and OpenCode. It sees every session, starts new work and tells you what matters. Type to it, or talk to it hands-free while you ride, run or walk.
curl -fsSL https://openmimir.com/install.sh | bash
You never pick an agent, a session or an id. Mimir knows where the work lives.
Sends follow-ups into the sessions you already have open and starts new ones headless. Risky tools stay blocked until you say so.
Picks up any Codex thread, including last week's, and reports back when it finishes.
Talks to your OpenCode service, so every message Mimir sends shows up live in the app.
“How's the billing refactor going?” or “pick up the review from last week” just work. It searches every agent's full history.
It listens while it talks and you can cut in. Live captions show what it heard, and it reads the plan back before starting work.
Trainer mode turns an iPad on the handlebars into a big, glanceable screen with huge approve buttons. Nothing to scroll.
Pushes, deploys and rm -rf wait for your approval, and voice can never approve them. It
all runs on your Mac.
Run the one-liner above. It puts mimir in ~/.openmimir/bin.
mimir init asks for an OpenAI key and where your projects live.
Run mimir, open localhost:4747 and press ⌘⇧Space.
OpenMimir is free. You bring your own keys, so you pay providers directly: voice uses OpenAI's
gpt-live-1 at about $0.05 per minute of conversation (roughly $3 an hour), plus the model
that runs the conversation and whatever your agents use.
Claude Code, Codex and OpenCode. You only need the ones you already use; Mimir finds them on its own.
Yes. Browsers only allow the microphone over HTTPS, so put Tailscale Serve (or a Cloudflare Tunnel) in
front of it, run mimir pair and open the link on the device once.
Your conversation and settings stay in ~/.openmimir on your Mac. Audio goes from your browser
straight to OpenAI's voice model; requests go to the model provider you choose. There is no telemetry.
Not yet. Desk and indoor trainer work today. Phone calls, a screenless mode for rides and runs, and a desktop app are next.
macOS first. The code is portable, and other platforms will follow.
curl -fsSL https://openmimir.com/install.sh | bash