ajsahni/agents-office

Agents Office v3 (Beta) – What it is

Agents Office is a locally‑run, visual “office” that lets you harness Claude (Anthropic’s LLM) as a team of 35 AI agents. The UI shows six department pods (e.g., Sales, Finance, Marketing) arranged in an isometric 3‑D office. You type a natural‑language task, the system routes it to the appropriate agent (or a team of agents), the agent runs Claude with the tools you have connected to Claude Code (Gmail, Slack, Notion, Chrome, etc.), and the result is saved back into a personal “Brain” – a folder of notes that the agents can read and write.

Core ideas

• Your Claude login – The software works with Claude Code (the CLI that authenticates to your Anthropic account) or a raw ANTHROPIC_API_KEY. All LLM calls are made from your machine, so no external server stores your data.
• Agents as roles – Each desk in the office is an agent with a fixed role (e.g., “Newsletter Agent”). Their behavior is defined by a short brief and optional skill files you can edit. Briefs give tone and constraints; skills are Markdown‑based SOPs that the agent can follow step‑by‑step.
• Task routing – A top‑bar lets you pick a department, type a task, and press Add. Claude first decides which agent should handle it, then runs that agent’s Claude session.
• Tools & connectors – The agents can call any service you have linked to Claude Code via claude mcp add (Gmail, Slack, Google Drive, Xero, etc.). A visual bar shows which connectors are available and which are actually used during a run.
• Routines & calendar – You can schedule recurring tasks (e.g., “every weekday at 8 am triage inbox”). Routines run automatically on the server’s clock, even when the UI is closed, and they respect a “needs my OK” flag for any action that would send, pay, or modify external data.
• Team mode – By pressing TEAM (or saying “as a team”), a lead agent splits a request into 2‑4 parallel sub‑tasks, each handled by a different desk. The lead recombines the pieces into a final deliverable.
• Claude in Chrome – With the official Claude‑in‑Chrome extension paired to your machine, agents can control a real Chrome window, read pages you’re logged into, fill forms, etc., under the same “read‑freely, act‑only‑when‑asked” rule.
• Model & effort selection – For each run you can pick Sonnet, Opus, or Fable, and set an effort level (Low‑Max). The UI displays which model and effort were used, and a usage gauge shows how much of your Anthropic plan has been consumed.

Getting started (as described in the README)

1. Install – Clone the repo, run ./setup (checks Node 20+, git, Claude login, installs deps, builds), then npm start.
2. Open – Visit http://localhost:4520. The right‑hand panel should read LIVE · CLAUDE.
3. Submit a task – Choose a department, type a plain‑English request, press Add. Claude picks an agent, runs the task, and the result appears both in the agent’s chat pane and as a note in your Brain folder.
4. Iterate – Use revise: … in the chat to ask the agent to modify its output. Corrections are stored as Markdown files that Claude can later turn into standing rules.
5. Customize – Edit office.agents.local.json (or use Claude in the repo) to rename agents, change their briefs, or assign different tools. Add skill folders under <brain>/Agents Office/skills/ to give agents reusable SOPs.

Who it’s for

• Power users of Claude who want a structured, visual way to run many LLM‑driven workflows without building their own orchestration layer.
• Solo entrepreneurs or small teams who need automated email triage, proposal drafting, bookkeeping, or content creation, and who already have Claude Code connectors set up.
• Developers who like to tinker – the whole stack is Node‑based, runs locally, and the agent roster/skills are plain JSON/Markdown, making it easy to extend.

Limitations (as of the beta)

• Runs only on macOS or Linux (Windows works via npm commands but the ./setup script is Bash‑only).
• All actions that affect external services require an explicit “OK” from you; otherwise the agent only reads.
• The UI is a demo‑style isometric office; it’s functional but not polished.
• The license is non‑commercial (PolyForm Noncommercial 1.0.0 plus Sahni.ai terms), so you can’t sell or embed the product in a commercial offering.

---

TL;DR: Agents Office is a locally hosted, 3‑D UI that turns Claude into a team of specialized AI agents, lets you route tasks, schedule routines, and hook into any Claude Code connector you have. It’s a concrete, open‑source implementation of an AI‑agent workforce for personal or internal use.