An agent-native
terminal for
your Mac.
Your AI agents work as a team — they talk to each other, span your machines, and answer when you call. You stop being the glue.
They talk to
each other.
No more relaying between windows. Two tools are the crux of agent communication — and every agent gets them on first launch:
converseAgents message each other directly — settle the API shape, the auth, the handoff. You stop being the messenger.delegateOne agent hands a whole task to another and gets the result back. You only weigh in to decide.
Same mesh, every session — local, or across machines over SSH.
Across every
machine.
An agent on your laptop and one on your server collaborate as if side by side — over plain SSH. No public ports. Sessions survive disconnects.
Stop typing.
Just talk.
Hit ⌘⇧V and think out loud — brainstorm, redirect, hand off a task, all mid-flow. Your agent talks back, and you can cut in any time.
- → full-duplex — interrupt and change course mid-sentence
- → no wake word, no always-on mic
You're not stuck
at your desk anymore.
Haku Mesh is the iPhone companion. Text or voice-call any session over your private Tailscale network — a real two-way conversation with the agents back on your Mac, from anywhere.
- → text any inbox · reply, redirect, escalate
- → call any agent — like FaceTime, but it answers
Nothing is
ever lost.
Your whole terminal history is searchable by intent. Ask for “that conversation about the auth change” and jump straight to the moment — then read the full transcript around it.
- → semantic search across every session you've run
- → find the moment, then open the whole transcript
- → on-device · private · nothing leaves your Mac
A browser preview,
built in.
Agents check their own work without leaving the terminal — navigate, screenshot, and run JS against the running app. SSH-aware: the server runs the code, your Mac renders the preview.
- → dev servers auto-detected and forwarded
- → open and read any file in the same pane
Better code search
for agents.
With grep, an agent that doesn't know your codebase's vocabulary has to guess a keyword, miss, and guess again — burning turns. Semantic search matches meaning: the agent searches for what it's actually trying to find and lands it on the first try. Same speed, far better aim.
Join the
waitlist.
// One email when access opens. That's it. No spam.