SkillAgentSearch skills...

agent-office

Desktop command center for AI coding agents: shared project memory, a context engine, a local MCP server and handoffs across Claude Code, Codex, Cursor, Gemini CLI and Copilot, shown as a live 3D virtual office.

Install / Use

claude mcp add arxdsilva -- npx -y github:arxdsilva/agent-office

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

75/100

Supported Platforms

Claude Code
Claude Desktop
Cursor
GitHub Copilot
Gemini CLI
OpenAI Codex

Our assessment of agent-office

agent-office scores 75/100 on our quality scale, 827th of 944 AI & Machine Learning skills we index.

Its MCP Server is 72 KB long, well organised into 40 sections with 7 code examples: long enough that it reads more like full documentation than a focused instruction file, which agents can find harder to follow.

It has 10 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
21/30
Structure
20/20
Description
15/15
Adoption
4/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated 3 days ago, so agent-office is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 97/100, with no cautions. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

agent-office compared with similar skills

All 4 of these similar skills score higher than agent-office; compare them before choosing.

SkillScoreStarsUpdatedFormat
agent-office (this skill)by arxdsilva75103d agoMCP Server
claude-memby thedotmack100118.5ktodayCLAUDE.md
Agent-Reachby Panniantong100113.3k12d agoCLAUDE.md
Understand-Anythingby Egonex-AI10093.3k1d agoCLAUDE.md
hyperframesby heygen-com10087.1ktodayCLAUDE.md

Frequently asked questions

How do I install agent-office?
Run claude mcp add arxdsilva -- npx -y github:arxdsilva/agent-office. The install tabs above show the steps for each supported agent.
Which AI agents does agent-office work with?
It is written for Claude Code, Claude Desktop, Cursor, GitHub Copilot, Gemini CLI and OpenAI Codex, as a MCP Server file. Other agents that read the same format can often use it too.
Is agent-office safe to use?
It is MIT-licensed and scores 97/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is agent-office still maintained?
The repository was last updated 3 days ago, so agent-office is actively maintained.

agent-office

agent-office. Your projects. Every agent. One shared memory.

An agent is temporary. A project is persistent. agent-office is a desktop app (macOS, Windows, Linux) that keeps your projects, their sessions, memory, decisions, tasks and handoffs in one encrypted local place, and shows the agents working on them in a 3D or 2D office.

The Project Dashboard: active agents, recent activity, completed work, tasks and project knowledge for one project

<table> <tr> <td width="50%"><img src="docs/screenshots/office-3d-projects.png" width="100%" alt="The 3D office with a project selected: its agents are highlighted in their wing"></td> <td width="50%"><img src="docs/screenshots/office-2d-projects.png" width="100%" alt="The 2D RPG office with the same project selected"></td> </tr> <tr> <td>The 3D office, focused on one project.</td> <td>The same project in the 2D RPG view.</td> </tr> </table>

Why agent-office

Coding agents start fresh every time. You re-explain the same decisions, a session ends and its context is gone, and a second tool knows nothing about what the first one did.

  • Projects outlive agents. A project holds its repositories, sessions, memory, decisions, tasks and handoffs. All of it stays readable when every agent is offline and after a restart.
  • Memory you control. Knowledge is kept as reviewable items with provenance. Agents can propose memory; only you approve it.
  • Any provider, one place. Claude Code, Codex, Cursor, Gemini, Copilot and any HTTP agent share the same projects. You can hand a task from one provider to another with a reviewed bundle.
  • Your standards, every agent. A harness turns your AGENTS.md, CLAUDE.md and Cursor rules into reusable, versioned engineering standards that managed sessions receive, that supported runtimes enforce, and that validation gates verify. The Compliance Inspector shows which rules apply, how each one is enforced, and whether the work passed.
  • Local and encrypted. Project data lives in an encrypted vault on your machine. Provider transcripts are referenced, not copied.
  • The office is still here. Agents are shown as workers in a 3D (or 2D) office, grouped by provider and conversation topic, so you can see at a glance who is working, waiting for you, or stuck.

Projects-first workspace

The app opens on the Overview once a project exists; the startView setting can pick the Overview or the Office instead.

  • Overview: every project, the agents waiting for you, recent completions and decisions, pending approvals, unfinished handoffs, unassigned sessions, and whether the vault is encrypted.
  • Project Dashboard: one project: active agents, recent activity, completed work, tasks, and knowledge (memory by kind). Actions start an agent, open memory or sessions, or prepare a handoff.
  • Sidebar: Overview, Office, Projects, Sessions, Memory, Harnesses, and the project list with counts.
  • Command palette: Cmd+K (Ctrl+K on Windows and Linux) searches projects, sessions and memory, and runs actions.
<table> <tr> <td width="50%"><img src="docs/screenshots/overview.png" width="100%" alt="The Global Overview: project cards with active agents, pending reviews and recent activity"></td> <td width="50%"><img src="docs/screenshots/palette.png" width="100%" alt="The command palette (Cmd+K) listing projects, sessions and actions"></td> </tr> <tr> <td>The Overview across all projects.</td> <td>The command palette.</td> </tr> </table>

Projects are found from repositories, not folder names. Several repositories can belong to one project, and a session whose repository matches nothing stays unassigned until you decide. Details: docs/projects.md.

Shared Memory and the Context Engine

Memory items have a scope (global, project, session, and an experimental workspace), a kind (instruction, constraint, decision, fact, note) and a state.

  • Review and approval. New items, including ones an agent proposes, start as pending. You approve, reject, edit, merge or supersede them. Edits keep earlier versions.
  • Provenance. Each item records where it came from: a Session Vault message, an imported note, or an agent proposal (shown with an "agent-proposed" badge).
  • Project MCP server. Managed sessions can retrieve approved memory through a local MCP server with the tools search_memory, get_project_context, list_decisions and propose_memory. Each session gets its own scoped token, revoked when the session ends. Results are labelled as reference data, not instructions.
  • Context Engine. It ranks the approved items a session may read (kind, scope, relevance to the objective, recency, pinned first) into a bundle with a size budget. Every item says why it was included or left out.
  • Context Inspector. For an agent it separates what is available (what the project policy exposes), what was retrieved (real MCP calls), and what was delivered (bundles agent-office put into a first message or handoff). It never claims what a model used internally.
<table> <tr> <td width="50%"><img src="docs/screenshots/project-memory.png" width="100%" alt="Project memory: items by kind and state, with approve, reject and edit actions and provenance"></td> <td width="50%"><img src="docs/screenshots/context-inspector.png" width="100%" alt="The Context Inspector: available, retrieved and delivered context for one agent"></td> </tr> <tr> <td>Memory review: pending items wait for your approval.</td> <td>The Context Inspector: available, retrieved, delivered.</td> </tr> </table>

Harness Builder — make your agents follow your flows and standards

A harness is a reusable, versioned set of rules and, optionally, a flow: instructions, runtime restrictions (deny or ask before a tool, path or command), validation gates (make check, go test ./...), ordered steps with checkpoints, and completion criteria. It describes any kind of work: a content pipeline (research → outline you approve → draft → review → publish), a research brief, a support reply, a release checklist, an incident write-up, as much as a coding task. Attach one or several to a project from the Harnesses sidebar pane (drag a card onto a project, or onto the project's Applied column; drag to reorder precedence) and every managed session on it receives the applicable rules and the current step, whatever the provider.

  • Build flows, not just rules. Steps have a goal, instructions, and exit checks: your approval, a passing gate, a file that must exist, or the agent reporting the step done (through the project MCP tool report_step, scoped to its own session). Step-scoped restrictions block actions that belong to a later step where the runtime can intercept. The Inspector shows a stepper with Approve, Mark done, Skip and Re-open; approvals also appear in the Overview.

  • Questions reach you. When an agent asks a multiple-choice question (AskUserQuestion), Agent Details shows the question card with its options; managed sessions are answered from there, observed ones show it read-only until you answer in the terminal.

  • Import your existing rules. Import rules from project scans the project's repositories for AGENTS.md, CLAUDE.md, .claude/rules/**, Cursor rules (.cursor/rules/*.mdc, .cursorrules), GEMINI.md and Copilot instructions, shows the candidates with duplicates and conflicts marked, and creates a draft with provenance (format, repo-relative path, hash). Nothing found is executed; provider files are never written. Publishing is your action.

  • One harness, multiple agents. The same project harness applies to Claude Code and Codex sessions (and is delivered as instructions to ACP providers such as OpenCode or Cursor), so you do not maintain separate standards per provider. A published version is immutable; running sessions keep the snapshot they started with.

  • Honest enforcement. Each rule is shown as Enforced by runtime (a trusted boundary blocks the action before it happens: Claude's PreToolUse hook, Codex's approval channel, a sandbox), Partially enforced, Instruction applied, or Not supported here. Shell command rules are always at most partial without an OS sandbox. A strict harness blocks a launch when a required rule cannot be enforced; an advisory one tells you and lets you proceed.

  • Verifiable completion. Gates run through agent-office's own trusted runner (no shell, scrubbed environment, redacted output). A session is Harness verified only when every required gate of the current harness version passed after the agent's last change; the provider's own state is never overwritten. Failures are fed back to the agent within a retry budget.

  • Compliance Inspector. Agent Details → Harness lists the applied harnesses, the coverage per capability, how many actions were allowed, blocked or asked, why each one was blocked, and the gate results. Minimal mode shows a compact badge per agent and subagent.

Shared Memory stays knowledge; a harness is policy. Agents have no tool to read or change harnesses, so they cannot weaken their own rules. Details and the provider coverage table: docs/harness.md.

Sessions and Handoffs

Every agent session becomes a session record that stays after the agent leaves. Its origin is managed (started in agent-office), observed (started elsewhere) or imported (past Claude Code history, experimental).

Automatic import into memory (experimental, off by default): turn on Settings → Import Claude sessions into memory automatically and confirm once. From then on, past and new Claude Code sessions are imported in the background, and each session's summary (its title, last replies, tools and duration; no model calls) becomes approved memory in its project, badged auto-imported, so your agents can retrieve it. Sessions not yet in a project stay private until you assign them. Memory → Remove imported memory takes all of it back in one step. Agent proposals still wait for your review.

  • Session Vault: browse sessions by project, provider and state, read agent-office's chat for a session, and open an old Claude Code session by reading its own transcript on demand.
  • Search: deep search across sessions and memory.
  • Save as memory: turn any Session Vault message into a memory draft that keeps the session and message as provenance.
  • Handoff: pick a session and a target provider, set an objective, review the context bundle that will be sent, and launch a managed session with it. The bundle is what you reviewed, nothing else. Handoffs need the target provider to have a managed runtime.
  • Tasks: you create tasks and confirm them. Each shows an evidence level (agent-reported, user-confirmed). Agents never write tasks.
<table> <tr> <td width="50%"><img src="docs/screenshots/session-vault.png" width="100%" alt="The Session Vault: a session's messages with a Save as memory action"></td> <td width="50%"><img src="docs/screenshots/handoff.png" width="100%" alt="The handoff dialog: objective, target provider and the context bundle to review"></td> </tr> <tr> <td>The Session Vault, with Save as memory.</td> <td>A cross-provider handoff with a reviewed bundle.</td> </tr> </table>

2D and 3D: projects in the office

The office is not a separate product: it shows the same projects. Both renderers use the same commands and panels.

  • Project Hub: the atrium room with the Project Directory kiosk. Walk up to it (or use the reception desk) to pick a project; the office focuses its agents and flies to its main wing.
  • Context Hub core: a glowing column ringed by six consoles. It is the project MCP station; it pulses only when a real memory retrieval happens.
  • Handoff Station: a pad in the plaz

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars10
CategoryAI
Updated3d ago
Forks2

Languages

TypeScript

Trust signals

97/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 info
agent-office — MCP Server: Install & Safety Check | SkillAgent