$ man agent-interface

Agents work
U stay in flow

A native workspace for a team of agents that keep working, coordinate with each other and keep what they make. Follow the progress, discuss the results and stay in the decisions.

macOS · Linux · Windows (through WSL) · open source · MIT licensed

← dark light →

Illustration with neutral example data
Drag across to compare themes; the page follows the switch at the top.

a responsibility, not just a prompt

Keep the work moving.
Keep your place in it.

In the example project acme-web, web is redesigning checkout, api is adding its endpoint, and docs is writing the release notes. Here’s your part.

  1. Give an agent a responsibility

    Ask web to redesign checkout in acme-web, keep you informed, and bring the confirmation wording back for your review.

  2. See the plan and progress

    Follow web’s current plan and the step waiting for your wording decision. Read api’s endpoint progress in Recent updates.

  3. Discuss the work where it lives

    Open the Checkout redesign plan Document and ask web to clarify the confirmation. Your message points to the revision you are reading.

  4. Follow the resulting changes

    Read web’s reply and the linked new Document revision in the thread. The version you commented on stays intact.

Illustration · example data

plans · artifacts · conversations

Talk to the work.
The agent who made it answers.

See what’s moving, point to what needs attention, and discuss it with the agent right beside the work.

  1. 01

    Waiting on you

    Answer the newest blockers beside the report that raised them.

  2. 02

    Current plans

    See each agent’s done, now, waiting and ahead steps; discuss a step.

  3. 03

    Recent updates

    Read progress from each agent and see its next actions.

  4. 04

    Responses

    Return to a saved question, answer and action, then follow up.

  5. 05

    Documents

    Discuss an exact revision; follow the author’s linked correction.

  6. 06

    Sheets

    Select cells and discuss them; agents update the workbook. Copy or export.

Illustration · example data

Conversations keep their record references. Messages never change the authored record.

one workspace · a team of agents

Different agents.
Shared bearings.

Run your team from the same window: Claude Code, Codex and OpenCode, with shared context and coordination you can follow.

  1. 01

    Policies

    Read the policies that reach your agents and discuss them in context.

  2. 02

    Skills

    Read shared skills, see who they reach and give the maintainer feedback.

  3. 03

    Team

    See who is working, idle or waiting for you, with their runtime and task.

  4. 04

    Activity

    Follow messages, safe queued delivery and issue handoffs between agents.

  5. 05

    Usage

    Inspect the tokens each agent and runtime reports.

Illustration · example data

Your agent CLIs work under their own permission settings. Messages never change the authored record.

[how it works behind the scenes →]

And underneath, a real terminal.

workspaces

One workspace per project. Everything in it, saved.

Tabs, the split panes inside them, and whatever runs in each pane: shells, servers, editors, agents. Switch projects and each one comes back exactly as you left it.

  1. 01Open a terminal⌘T
  2. 02Split right⌘D
  3. 03Split below⌘⇧D
  4. 04Switch workspace⌥⌘→
storefront3 tabs
research2 tabs
Jump to tab, pane⌘K
TABS
1api
2web
3db
1notes
2papers
Terminal~/storefront1
❯ claude
> Add rate limiting to the API
⏺ Read(src/routes/mod.rs)
⎿ Read 142 lines
⏺ Write(src/middleware/limit.rs)
⎿ Wrote 64 lines
⏺ Bash(cargo test)
⎿ 48 passed; 0 failed
✶ Writing tests…
Terminal~/storefront2
❯ cargo watch -x run
INFO listening on :8080
INFO GET /health 200 0.4ms
INFO GET /v1/items 200 3.1ms
WARN slow query 212ms
Terminal~/storefront3
❯ git status --short
M src/routes/mod.rs
?? src/middleware/limit.rs
❯
Terminal~/research1
❯ codex
› Summarise papers/ by method
• Ran ls papers | wc -l
└ 37
• Grouping by method…
Terminal~/research2
❯ python3 plot.py --all
loaded 37 papers
wrote figures/methods.png
❯

what you get

A real terminal, with a desktop around it.

Terminal~/storefront
❯ npm run dev
ready on :3000
Terminal~/storefront
❯ htop
⌘D · ⌘⇧D · drag · ⌘⇧↩

Split, resize, expand

Split right or below, drag a divider to resize, and give one pane the whole tab when you need room.

builder
⏺ Update(routes/mod.rs)
⎿ 2 additions
⏺ Bash(cargo test)
⎿ 48 passed
✶ Writing tests…
reviewer
• Ran cargo clippy
└ no warnings
Key on the API
token instead?
› 1. Yes
claude · codex · any CLI

Any agent, any program

Claude Code, Codex, a dev server, an editor: if it runs in a terminal, it runs here, side by side.

editorserverlogsapi-server
TMUX SESSIONS
1api-server3 windows
2scratch1 window
⌘⇧A attach

Bring the sessions you have

Attach any tmux session into a pane and click between its windows. Nothing disconnects, and nothing is created until you choose.

ser⌘K
server · cargo watchpane · api
servicestab · infra
⌘K jump

Jump to anything

Tabs, panes and commands across every workspace, a few letters away.

window closed
tmux claude cargo watch zsh
tmux underneath

Sessions outlive the window

Close a pane, detach a session or quit the app. tmux keeps every process running, and reopening puts it all back. How it works →

❯ cargo test 2>&1 | tail -4
test limit::allows_burst ... ok
test limit::rejects_flood ... ok
test routes::health_ok ... ok
test result: ok. 48 passed
❯
Copied
GPU-drawn · Ghostty's core

A real terminal, drawn natively

Scrollback, selection, copy and bracketed paste, with mouse support for the programs that want it. Never a web view.

Rename tab⌘R
Move up
Move down
Move to workspace
drag · rename · move

Tabs that move with you

Drag tabs to reorder, rename with a double-click, or move a whole layout to another workspace.

Skillsevery skill an agent can load
Shared skills
rust3 agents
cargo-workflowalso review
api-security2 by name
skills · team · activity · usage

Native views, when you want them

Skills, Team, Activity and Usage open as native panes beside your terminals. See how it works →

ThemeDarkLight
Text size− 13 +
Pane headersAlwaysOn hoverNever
Split right⌘DChange
⌘, settings · ⌘N new window

Make it yours

Light or dark, your text size, pane headers always or on hover, every shortcut editable, and as many windows as you like.

set up agent interface

One command.
Then it shows you around.

One command installs Agent Interface and everything it runs on. Open the app, let onboarding check your computer, and meet your first orchestration agent.

curl -fsSL https://raw.githubusercontent.com/eandualem/agent-interface/main/install.sh | sh

macOS · Linux · Windows (through WSL) · Same setup on Windows: run inside WSL. · install guide ↗

  1. 01

    Check

    tmux, Agent Backbone, Git and the GitHub CLI, checked on this computer.

  2. 02

    Agent tools

    Claude Code, Codex, Gemini CLI or OpenCode. You sign in inside each tool.

  3. 03

    Your agent

    Pick a tool and folder. Your coordinating agent starts, introduces itself and opens in its own terminal.

  4. 04

    Shortcuts

    The keys you’ll use most. Change any of them in Settings.

Illustration · example data

support

Get your team
off the ground.

Hands-on help with Agent Interface setup, agent responsibilities, shared skills and policies, and the workflows that connect them.

[talk to the people who build it →]