Before and after
Four agents, two ways to watch them
Inbox✓ All clear · back to your work
- A single sessionNothing to orchestrate. The terminal is enough.
- Two or moreYou need a list of who is waiting, and a separate checkout if they edit the same repo.
| Several terminals | With doki | |
|---|---|---|
| See who is asking | Check every window | Inbox sorted by attention |
| Reply | Switch apps and type | 1, 2, or 3, or a message |
| Finished work nobody saw | It sits in its tab | A “done” state on the same list |
| Same files | Conflicts | Worktree, port, and .env per agent |
| One builds, one reviews | You coordinate by hand | A combo until approved or stuck |
There is no config file for “adding” agents
doki reads transcripts (Claude in ~/.claude, Codex in ~/.codex, Droid in ~/.factory, Kiro in ~/.kiro), Cursor’s local database, and the processes of the rest. You open the agent as usual. Within a couple of minutes it is on the list, with the project next to the name.
Claude’s real time is the only optional step: in Connections, “Enable real time (hooks)” writes to ~/.claude/settings.json after saving a backup. Without that, Claude still shows up, with less immediacy.
Attend only to what is stuck
An agent that is working does not need a visit. One that finished needs a look at the diff, not a conversation. One that is asking needs an answer now. The Inbox separates those three, and an error is marked on its own.
The shortcuts are in the Inbox guide. If you would rather not read the list, the voice router sends what you say to the agent already on that task, or opens a new one.
Isolate the repo so parallel means parallel
Two agents in the same directory fight over files, the app’s port, and .env. doki can give them a worktree: a separate branch, a reserved port, a copy of the environment. The operational detail is in worktrees.
That also leaves a clean path to merge, open a PR with gh, or discard. What the agents did can be exported to CSV or JSON.
One builds and another reviews
A combo chains two agents until the result is approved or they get stuck. The reviewer is not a human skimming: doki can pass the diff and return a verdict with risks and secrets, then send those findings back to the agent to fix.
It works for Claude against Codex, or Droid against Cursor. The brands matter less than each one having its own session and checkout.
Parallel can also mean another computer
A Mac mini or a Linux box can add sessions to the same Inbox, over SSH, with no ports opened. If the agent runs in tmux, typing is live. The laptop can sleep.
The phone joins the same circuit. It is not a second product: it is the Inbox when you are not in front of the Mac.
Frequently asked questions
How many agents does doki support?
As many as your Mac can run. The list includes Claude Code, Codex, Cursor (app, CLI, and Cloud), Gemini CLI, Grok, Factory Droid, Kiro, OpenCode, Amp, Aider, Goose, Qwen, Hermes, Crush, and Pi, among others.
Do I have to launch them from doki?
No. doki finds the ones you already started. You can also assign a new one from Talk or, for Cursor, through the API.
Do the conversations merge?
No. They share the attention queue. Each thread still belongs to the agent that started it.
Does this replace the agent?
No. Each model is billed where you already pay for it. doki is the Mac layer that steers them.