Paperclip explained, and compared with similar projects

Checked on 2026-10-29 against Paperclip's GitHub page, official docs and changelog, and the GitHub pages of the four comparison projects. FollowAgents has not run any of them hands-on. Stars, versions and features change quickly; where this page differs from the linked official pages, the official pages win.

Paperclip is an open source (MIT) orchestration platform for AI agents. Its own tagline: "If OpenClaw is an employee, Paperclip is the company." It does not teach agents how to think. It gives agents you already have (Claude Code, Codex, Cursor, OpenCode and more) the management layer of a company: an org chart, goals, tickets, budgets, approvals and an audit log. At the time of checking the repository showed about 108k stars and 18k forks, the most popular project of its kind.

This page explains what Paperclip is and is not, its core concepts, supported agents, setup and the risks to know first, then compares it with OpenCompany, ai-employees, Eigent and AI-company (AI Team OS). These five projects take very different shapes, so the goal is to help you decide which kind of tool matches your need, not to produce a ranking.

In one paragraph: what it is and is not

What it is: a management layer that runs on your own machine or server. A Node.js server plus a React dashboard, with an embedded PostgreSQL database by default, served at http://localhost:3100. You create a company, set goals, "hire" agents and assign them tickets. Agents wake on "heartbeats" to work, and every action is recorded.

What the project says it is not (its own statements):

• Not a chatbot: conversations stay attached to tasks and outputs.
• Not an agent framework: it does not dictate how agents are built; agents bring their own prompts, models and runtimes.
• Not just a workflow builder: routines operate within roles, goals, budgets and governance.
• Not a prompt manager, and not limited to one agent or to code review.

Its bar for joining is one line: "If it can receive a heartbeat, it's hired."

Core concepts

Two things to note:

• Budget stops are not instant. The docs say enforcement relies on recorded spend, so in-flight work can delay a stop. Don't treat it as a precise circuit breaker.
• Companies can be exported and imported. Organizations, agents and skills (and optionally projects, routines, tasks and attachments) can be packaged. Secret values are left out, but the README warns plain environment values and local paths can remain, so review a package before sharing it.

Which agents it supports (adapters)

The official adapters page lists 13 built-in adapters, most of which drive an existing agent through its command line on the Paperclip host:

External adapter packages can extend the list. Details: most local adapters need an absolute cwd; secret references are preferred over putting sensitive values in env; and the docs note Cursor and native Grok don't provide authoritative per-run cost totals. The README roadmap also lists cloud and sandbox runtimes such as e2b, Cloudflare, Daytona and Modal.

Setup and first run

Requirements: Node.js 24.11 or newer; pnpm 9.15 or newer for a source install. To sign in with a Claude or Codex subscription, the Paperclip host also needs Python 3 and the matching provider CLI.

One-line start (the README's default):

npx paperclipai@latest onboard --yes

To reach it from other devices on your LAN or Tailscale, use the authenticated private mode:

npx paperclipai@latest onboard --yes --bind lan
npx paperclipai@latest onboard --yes --bind tailnet

Just want to look around? test-drive creates an isolated instance with a pre-made CEO agent:

ANTHROPIC_API_KEY=... npx paperclipai test-drive
OPENAI_API_KEY=... npx paperclipai test-drive --harness codex

From source:

git clone https://github.com/paperclipai/paperclip.git
cd paperclip
pnpm install
pnpm dev

The official quick start then has three steps: create a company and give it a goal, hire a first agent (a CEO in idle status), and watch the heartbeat — the CEO submits a strategy approval first, and after you approve it the first tasks appear on the board. See the official five-minute path.

Three deployment modes (official deploy overview):

Know this before you use it

• It costs real money. The quick start says agents make paid API calls: roughly $5–20 to experiment and $20–100 a month for an active company (the project's estimate). Set per-agent and company budgets before enabling heartbeats.
• The default local mode has no login. local_trusted is for one person on their own computer. To let others in, switch to an authenticated mode, and don't expose it straight to the internet.
• Agents run with your machine's permissions (our inference). Local adapters drive the Claude Code, Codex and other CLIs installed on the host, so what they can read, write and execute depends on each tool's own sandbox and approval settings. The adapters page has no dedicated permission warning, so tighten those settings first; see our Codex guide and OpenCode guide.
• Telemetry is on by default. The README says it is anonymous and excludes personal information, issue content, prompts, file paths and secrets. Turn it off with PAPERCLIP_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1; it is off automatically when CI=true.
• It moves fast and breaks things. Several recent changelog entries list breaking changes (for example v2026.1012.0 changed the workspace restore-lock protocol and how stored tool grants are enforced). Read the changelog before upgrading.
• About 3,400 open issues and 3,200 open PRs at the time of checking. High popularity, but also a sign of rapid change; validate in an isolated environment before relying on it.

Compared with OpenCompany, ai-employees, Eigent and AI-company

The short version: all five use "company", "employee" or "team" language, but they solve problems at different layers.

OpenCompany: the employee as a product

OpenCompany (MIT) pitches "self-improving AI employees running on your own computer". You write a job description in plain language to "hire" one; each employee has a workspace, routines, skills and memory, with a lead agent coordinating specialists. Work can arrive by chat, email, WhatsApp, Telegram, Discord, schedules or webhooks, and there is a browser you can take over. The stack is a Python backend, a React interface and Temporal for durable background work. The big difference from Paperclip: it ships its own runtime and is about one employee finishing a job, whereas Paperclip ships none and is about running an organization of agents. Its page states that installers are not yet code-signed, that background work needs the app running and the computer awake, and that "self-improvement" means saved memory, clearer instructions and new skills — the model is not retrained.

ai-employees: just prompts, but very light

ai-employees (MIT) is essentially plain text: one folder per role with ROLE.md, SCHEDULE.md and per-routine SKILL.md files, plus an npm installer and a Claude Code plugin. npx ai-employees hire <role> drops a kit into your project. It has no server, no dashboard and no budget mechanism; your existing agent (mainly Claude Code) does the work, using your signed-in browser. The project says sends, posts, submissions and spending are staged for your approval. It suits an individual who wants a set of marketing, sales or SEO routines inside Claude Code quickly. Caveats: the README is inconsistent about how many agents it supports (10 vs 12), its self-repair claims are project statements we have not verified, and the author also promotes a paid community while stating the core repo is free.

Eigent: a desktop app for cowork-style tasks

Eigent (Apache-2.0) calls itself an open source cowork desktop app and a local, free alternative to Claude Cowork and Codex. It is built on the CAMEL-AI multi-agent framework, with an Electron and React front end. It offers a single-agent mode and a Workforce mode where specialised agents run in parallel, with MCP, skills, and built-in browser and terminal toolkits; models can be cloud APIs or local inference through vLLM, Ollama or LM Studio. It targets "I have a multi-step job, let a group of agents finish it", not long-running operation of a company with budgets and approvals. Points to note: its quick-start path requires an Eigent cloud account while a fully local setup follows a separate server guide, SSO and similar features go through sales, and its roadmap still lists unfinished Workforce items. Our catalog covers CAMEL, the framework it builds on.

AI-company (AI Team OS): a coordination layer for coding agents

AI-company (named AI Team OS, MIT) is conceptually closest to Paperclip but the smallest in scale. It is also a "team operating system for agents", but only for coding agents such as Claude Code and Codex. It doesn't run agents itself; it lets them share a task wall (with ownership and dependencies), project memory, team messages, 8 meeting templates and a 4-round debate mode, with 25 role templates and a React dashboard. It installs as a Claude Code plugin or by script and stores data in local SQLite. Its own page lists the gaps: Codex token and cost attribution is incomplete, an idle Codex session isn't woken by messages, multi-tenant isolation and PostgreSQL are still planned, and the tool and template counts in its README and its About text don't match. It reads as an experimental layer for an individual or small team rather than a governance platform like Paperclip.

How to choose

• You manage several different agents and need budgets, approvals, audit and multiple users: Paperclip. It is the only one of the five where governance is the headline.
• You want an out-of-the-box "AI employee" with its own runtime, reachable through chat and email: OpenCompany.
• You only want ready-made role routines inside Claude Code, with no server: ai-employees.
• You want a desktop multi-agent workspace and lean toward local models: Eigent.
• You only want several agents in Claude Code / Codex to share tasks and memory: AI-company, after reading its stated limits.

They are not mutually exclusive: Paperclip drives Claude Code through an adapter, while ai-employees' role prompts are themselves Claude Code skill files. Decide first whether you need a runtime, a prompt kit or a governance layer.

Who it fits, and who it doesn't

A good fit: individuals or small teams already running several coding or office agents who are losing track of who is doing what, what it costs and whether anyone approved it; people who must keep agents for different projects or clients in separate "companies"; and people willing to try it in an isolated environment before handing it real authority.

Not yet a fit: someone who uses one agent for occasional code; anyone unwilling to carry API costs and keep an eye on runs; anyone who needs a stable interface (it moves fast and has breaking releases); and anyone who would expose the loopback mode to several people.

Recent releases

Per the official changelog, the latest at the time of checking was v2026.1023.0 (2026-10-23):

• v2026.1023.0: tasks and projects can be private and shared with chosen people; costs are split into estimated, unpriced and subscription spend versus API spend; fourteen apps were added to the connector catalog.
• v2026.1012.0: outside assistants such as Claude Code, Codex and OpenCode can sign in as you through MCP to read, create and delegate work; two breaking changes.
• v2026.1005.0: agents keep their own files across tasks, company skills can sync from GitHub folders, a live cloud browser and five experimental memory connectors arrived; @-mentioning an agent no longer starts a run.

On the roadmap, the README marks memory and knowledge, Agent Chat and cloud deployments as in progress, and a desktop app, bring-your-own ticket system (Asana, Linear, Jira) and self-organization as planned.

FAQ

Is Paperclip free? The software is MIT-licensed and self-hostable. You pay for the models and agents you connect (API fees or their subscriptions). The website has a waitlist for Paperclip Cloud; we have not checked pricing.

How does it relate to OpenClaw? OpenClaw is one kind of agent Paperclip can connect, the "employee" in the project's metaphor. The OpenClaw Gateway adapter is currently API/import only. See our OpenClaw profile.

Does it replace Claude Code? No. Paperclip doesn't write code; it schedules Claude Code, Codex and others to do so. See our Claude Code profile.

Can I run it fully offline with local models? OpenCode's adapter supports provider and model routing, so a local model is plausible in theory, but we have not verified that path.

Has FollowAgents scored it? No. Paperclip is not in our catalog yet, so this page is a reference write-up, not a FARS score or a recommendation.

Official sources

• Paperclip on GitHub (README, licence, roadmap)
• Paperclip docs: five-minute path, adapters overview, deploy overview, changelog
• Compared projects: OpenCompany, ai-employees, Eigent, AI-company
• Related on this site: AI coding agents compared, Hermes Agent guide