Agent Office

English · 한국어 · Changelog

Bots and automations scattered across your AI clients — Codex, Claude Code, Hermes, scripts — managed from one screen. Each one still runs on its own app and your own accounts. In Agent Office you see their status, hear when one stops, change instructions and run them again. Their results come together in one feed.

Full-quality MP4 · how it is made

What's new

• v0.8.0 · Business profile · separate ports · several Offices' MCP on one PC
• v0.7.6 · Security fixes
• v0.7.4 · Aside recommended · one server alert per check

All releases

Nothing stops silently: a stopped run shows up in Needs you and, if you connected one, in your messenger. New recurring work takes one line and runs when it is due.

It is also a local MCP server, so Codex, Claude Code, Cursor, OpenCode, Hermes or another MCP client can hand it work.

The feed: where results come together

The more work you hand to AI apps, the more the results scatter. The morning briefing lands in Telegram, the cleaned-up table in a work folder, the email draft somewhere in a chat. With ten recurring tasks you spend time asking where yesterday's result went. The board tells you how each Work is doing, but what you actually read every day is the results.

So we built the feed. It puts the results of every Work in one stream, newest first, and shows each one in the shape it is used.

• One stream. Results Office saved, messages your server bots sent, the last answer of a conversation you attached, and results an AI app sent elsewhere all land in the same place.
• Shaped for use. Writing reads like an article, options sit side by side, an email draft waits for your check before it goes out, spending shows as a number and a chart. There are checklist, event and price-change cards too.
• Your turn comes first. A Work that needs a choice or a check before sending is pinned at the top, and you answer right there.
• Reply to a result. Send the next instruction — "shorter", "look again for this flight" — from under the result to the same Work.
• On your phone. Turn on notifications and new results reach your phone. What is running now and when the next results are due sit beside the stream.

Side by side	Check before sending
A number and a chart	Writing that reads like an article

A look inside

Work board — describe the work in one line and follow it by status. The sidebar holds Feed and Work; All, Needs you, Running and the other views are tabs inside Work.

Timeline — when each Work ran in the last 24 hours and how it ended.

Work detail — the app's own run as it happened, the completion check it reported, and the result files.

The Works and results on the board, timeline and feed are samples. The detail is a real run of a made-up task (summarise a three-month sales log into report.md) with Codex on a fresh install.

Supported environments

See Work operation for the execution loop and release verification for tested scope and remaining limits.

The default browser is a dedicated background Chromium. A VM is optional.
Websites are handled in the browser; clicking and typing in apps installed on the PC, outside the browser, is still experimental.

Get started

Ask your agent:

Install github.com/deepdivekr/agent-office and connect it over MCP. Use Agent Office for my browser and file work.

Or run this from Ubuntu or WSL Ubuntu, in any directory:

bash -o pipefail -c 'curl -fsSL https://raw.githubusercontent.com/deepdivekr/agent-office/v0.8.0/install.sh | bash'

You can review the installer first.
It prepares a dedicated Node.js, dependencies, and Chromium, then opens the Control Center.
If Chromium's system libraries are missing, the installer asks once at the terminal and runs Playwright's install-deps chromium with sudo, which installs official Ubuntu packages with apt. sudo reads your password; Agent Office never sees it. Nothing else runs with admin rights, and without a terminal the installer prints the command instead.

Connect in the Control Center

Connect your tools once, then hand work over from the board.

1. Clients — check installation and sign-in, then register MCP.
2. Execution — use the default browser or connect another executor.
3. AI — choose the app that runs your work (Codex or Claude Code) and its default model.
4. Delivery (optional) — save Telegram, Slack or Discord destinations. Results always stay in the feed.
5. First Work — enter instructions, a completion condition and destinations. Leave the condition blank and the AI writes it.

Jev is optional. Without it, the LLM and code handle decisions.
Website sign-in is requested when a Work needs it.

Aside and Neo need their own installation and running app. Check and register them in the Control Center; without them, Playwright is used. Browser setup

To reopen the Control Center, run this in the same Ubuntu environment:

~/.local/bin/agent-office connect

For manual MCP registration, use agent-office mcp. Windows apps use the WSL command shown in the Control Center.
First-run guide · MCP configuration

A second Office on the same computer

Keep personal and business Work apart with a profile. Each profile has its own data (~/.agent-office-profiles/<name>), a fixed port of its own (4601–4699) and its own MCP name (agent-office-<name>):

~/.local/bin/agent-office --profile business connect

The profile's name is shown next to the logo, and a Work in one Office cannot start or change Works in the other. Both run as your account, so a profile separates records and notices, not file access; for business data that other Works must not reach, use a separate Windows account or computer.

What can I ask it to do?

Every morning at 7:30, send me the three main news stories with sources on Telegram.

Compare Seoul–Osaka round-trip fares for the second week of November every day and tell me when prices drop.

Every hour, sort new customer emails and pick out the ones that need a reply.

Sort this month's card statement by category and compare it with last month.

Write an appointment request to the clinic. Show it to me before sending.

A Work is one such request, with its run history, completion checks and progress in one place.
The public acceptance set (checking an official site, saving public data as CSV, a form draft and two more) completed end to end without a click in between, with Codex as the subscription model, in 1.5–3 minutes each. Record

How it behaves by default:

• Ask first. It asks the conditions that decide the result before it starts. Uncheck it for a quick run.
• Runs to the result. After you allow AI use once, a Work you start runs to its result and keeps its own schedule. Submissions, payments and messages to other people still ask you every time.
• Drafts are not submitted. A form draft is filled in a private page that is closed afterwards.
• The browser stays out of your way. A background browser reads first; if a site refuses it and Aside is connected, only that read moves there.
• Gets faster with use. A Work that passed verification leaves its procedure behind, and a similar request repeats its reads without model turns. Saved procedures and remembered sources are listed under the board, where you can turn them off.
• Small MCP surface. agent-office mcp lists 10 Work tools (about 2k tokens). Every other tool stays callable by name; --all-tools lists all 114.

From request to result

1. Start with one request. The detail opens; the chosen AI analyzes the work, asks what matters, and runs it.
2. Follow the run record: the AI app's commands, files and messages, the sources it read, and why it is waiting.
3. Adjust a stage to pause, change instructions or continue. A new direction goes to the same session of the same app as you wrote it.
4. Read the result in the feed and the detail, with sources and files. It is not marked complete before its completion checks pass.

New Work keeps its results in the feed, and the messengers you pick receive them too. Delivery setup and limits

The nine built-in Pack families are starting points.

You do not need to write a Pack first. The LLM structures the request and reuses verified procedures.
Work you repeat often can be saved, once verified, as a named and versioned custom Pack. Custom Pack operation

Who runs the work?

Each Work runs on the AI app chosen when it was started (Codex or Claude Code, with a model and reasoning effort), using that app's own settings, skills and MCP servers.
That app keeps the Work: when its sign-in or allowance runs out, the Work waits, and Office never moves it to another app by itself.
To move a Work yourself, use Change next to the app name on the Work detail, between runs. Office reads the earlier runs for what the new app lacks (Codex's built-in image generation, an MCP server, a skill) and drafts a handoff direction that goes with the change, such as leaving only the pictures to codex exec. The next run starts a new session on the new app with the same Work folder and records.

• LLM: splits the work, handles unfamiliar situations, and replans.
• Jev (optional): short, typed decisions defined by the Pack.
• Code and executors: clicks, typing and file work, with result checks.
• Runtime: checkpoints, sessions and run records. Uncertain writes are not blindly replayed.

Bring in what already runs

Use Import for work that already runs elsewhere. Its runtime, schedule and messenger delivery stay as they are; importing alone never starts a duplicate or a new schedule.

• Hermes and remote OpenClaw keep running there; status and directions go through Office.
• Services on your server are watched read-only over SSH: systemd services, timers and Docker containers. Timer runs appear on the timeline, and messages the bots sent appear in the feed.
• AI conversations from the Claude Code or Codex desktop app or terminal become Works you can continue from Office while they are idle. Under WSL, conversations of the Windows apps can be attached too.
• Automations in another app: paste the copy prompt into that AI and turn its answer into a Work draft.

Work import · Remote management

Updates and limits

In AI settings, turn on Auto-update CLIs or press Update now. While the Control Center is open and no Work is running, Linux/WSL CLIs are checked every 24 hours; apps in use are left alone. Windows automatic updates are not supported yet.

To update Agent Office, finish active work, close MCP clients and the Control Center, and rerun the installer.
If the shared server is still running, follow the stop and reconnect guide.
The app lives in ~/.local/share/agent-office, settings and Work data in ~/.agent-office. Existing ~/.agent-driver data is reused in place.
Installations with private workflow code need a compatibility review first.

• Aside and Neo adapters are read-only for now. Form writes and guest-VM transports are not supported.
• People handle sign-in, CAPTCHAs, and approvals.
• Keep local connection tokens, API keys, and private workflow data out of shared files.

Release validation · AI settings · Browser routing · Memory and process lifecycle

Development and license

git clone https://github.com/deepdivekr/agent-office.git
cd agent-office
npm ci
npm test

Use Node.js 22.22.0 and npm 11.11.0. The default test suite excludes long soak tests.

README pictures are made with node scripts/docs/capture-readme-feed.mjs en|ko (board, timeline and feed with sample data). The Work detail is a real run of the sample sales-log task on a fresh install (agent-office connect --state-root <new folder>).

Apache License 2.0 · © 2026 SHIPIT LABS. Connected models, browsers, and services have their own licenses and terms.

Made by SHIPIT LABS · news and questions on X: @DeepDive_KR