Paperclip vs Claude Code

At a glance

What Claude Code does well

Claude Code is the agent. It reads your repository, edits files, runs commands and opens pull requests, and it does this from the terminal, an IDE, a desktop app, the web or the Claude mobile app.

Agent teams are its newest way to go wide. You set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, describe the team in a prompt and a lead session starts teammates. Each teammate has its own context window. They share a task list with dependencies, message each other through a mailbox and can run in split panes with tmux or iTerm2. Hooks such as TeammateIdle and TaskCompleted act as quality gates.

Subagents cover the smaller case. A subagent does one job in its own context and returns a summary to the caller, at a lower token cost than a teammate.

The rest of the toolkit is strong too. claude --worktree starts a session in an isolated git worktree. claude --bg starts a background agent that keeps running after you close the shell. Cloud sessions run on Anthropic-managed machines, start from a browser or a phone and can fix a pull request when CI fails. Cloud scheduled tasks run on a schedule or on a GitHub event.

Anthropic also documents the limits. A session has one team, teammates can’t start their own teammates, /resume doesn’t restore in-process teammates and a team uses many more tokens than one session. Headless -p mode doesn’t start teammates at all.

Where Paperclip differs

One space for the whole team

A Claude Code session, with or without an agent team, belongs to one person. Its context, history and approvals stay on that machine. Paperclip is multi-player. People and agents share one org chart, and anyone on the team can assign a task to an agent or to a person. Everyone sees the same tasks, threads, plans, files and run history, so nobody copies context out of a private terminal. Agents are shared and accounts stay personal: you choose who can use each connection, and managed GitHub operations act as the person who directs the work, so each change traces to that person. Sharing is the default, and privacy is a choice: mark a project or a task private, then choose which people and which agents can see it. Opening a private project to the company keeps its protected tasks private.

A task queue instead of a session

In Claude Code the unit of work is a session. In Paperclip the unit of work is a task with one owner and a thread. You create tasks, assign them to agents and leave. Paperclip wakes an agent when it gets a task, when someone comments and on a schedule. The thread keeps the plan, the diff and the results after the run ends, so the work survives a reboot or a closed laptop.

People sign off before work ships

Claude Code’s GitHub code review posts comments and a check, but it doesn’t approve or reject a pull request. Paperclip can give a task a Review stage and an Approval stage. When the agent marks the task done, Paperclip moves it to In Review. The task closes only after approval, and “changes requested” sends it back to the agent. A person or another agent can be the reviewer.

Budgets with a hard stop

Claude Code can cap one print-mode run with --max-budget-usd, and team plans can set spend limits for an organization. Paperclip sets a budget on the company, an agent or a project. Each budget warns at a threshold, 80% by default, and pauses the agent at 100%. A loop can’t use your whole quota overnight.

Any agent, from any vendor

Claude Code runs Claude models. Paperclip runs Claude Code, Codex, Gemini CLI, Cursor, OpenCode, Hermes and more, each with the login you already have. You can give one task to Claude Code and the review of that task to Codex.

Apps with a rule for every action

Paperclip connects to 40+ apps and model providers, including GitHub, Linear, Notion, Slack and Google Workspace, and to any MCP server by URL. You choose who can use each connection and which agents can reach it. Every action has its own rule: Off, Ask first or Allowed. Ask first turns the call into a decision for a person. Claude Code reaches apps through MCP servers with allow, ask and deny permission rules for each tool. Its skills live in files in your project or home folder, shared through git or plugins, and cloud scheduled tasks cover the schedule case.

Routines and shared skills

A routine is repeatable work with a schedule or a webhook trigger. Each run creates a task you can follow, and the routine keeps a history. Skills are shared across the whole company. Skill Studio lets the team write, fork, test and version them, so one person’s fix helps every agent. Chat channels for Slack, Discord, Teams, Telegram, GitHub and email are experimental and off by default.

Use them together

Paperclip runs Claude Code as one of its agents, non-interactively through ACP. On the agent’s Connect step, you sign in to your Claude plan or add an Anthropic API key. Paperclip keeps that sign-in separate from your personal Claude Code login. Inside a Paperclip task, Claude Code still uses its own tools, hooks and subagents. Whether agent teams start inside a Paperclip run is not something we have tested, so this page doesn’t claim it.

• A teammate assigns the task
• Anyone comments, the agent wakes
• The reviewer approves

1. 1TaskOne owner. A thread for the plan and results.
2. 2Claude Code runsWoken by the task, a comment or a schedule. Optional git worktree.
3. 3Pull requestThe agent opens it with its own git and gh.
4. 4Review stageA named reviewer, a person or an agent, approves or requests changes.
5. 5ApprovedThe task closes only after approval.

When to choose Claude Code

• You want several views of one problem right now, such as a design review or a bug with three possible causes.
• You work in one repository and watch the work as it happens.
• You want a native desktop app or a terminal and nothing else to run.
• Your whole team is on Claude and you don’t need to mix vendors.

When to choose Paperclip

• Your whole team shares the agents, with the same tasks, threads and history, and sensitive work stays private to the people and agents you choose.
• You have more tasks than you can watch, and you want to come back to finished work.
• You want a person or an agent to sign off on every change before it ships.
• You need a spending limit for each agent or project with a real stop.
• You want Claude Code to build and Codex to review, or the other way round.
• The same agents also handle support, marketing or operations tasks.

FAQ

Does Paperclip replace Claude Code?

No. Paperclip runs Claude Code and other agents. It adds tasks, review stages, budgets and schedules around them.

What is the difference between Claude Code subagents and agent teams?

A subagent works for the main session and returns a summary to it. Teammates in an agent team are separate sessions that share a task list and message each other directly.

Do Claude Code agent teams work inside Paperclip?

We don’t know yet, so we don’t claim it. Paperclip runs Claude Code non-interactively, and Anthropic’s docs say headless mode doesn’t start teammates.

Can several people use the same agents in Paperclip?

Yes. Paperclip is multi-player. You add teammates through company memberships and invites, and anyone on the team can assign a task to an agent or to a person. Everyone sees the same tasks, threads and run history.

Are Claude Code agent teams free?

They are part of Claude Code, so they use your Claude plan or API usage. Anthropic’s docs say a team uses many more tokens than a single session.

Do I need an API key to run Claude Code in Paperclip?

No. You can sign in to your Claude plan on the agent’s Connect step inside Paperclip. Paperclip keeps that sign-in separate from your personal Claude Code login. An API key is the other option.

Sources

• Anthropic, Agent teams, checked November 10, 2026
• Anthropic, Subagents, checked November 10, 2026
• Anthropic, CLI reference (--worktree, --bg, --max-budget-usd), checked November 10, 2026
• Anthropic, Code review, Scheduled tasks, Costs and Overview, checked November 10, 2026
• GitHub, anthropics/claude-code, 155,196 stars, license “All rights reserved”, checked November 10, 2026
• Paperclip, README and docs, 112,884 stars, checked November 10, 2026
• Paperclip docs, Anthropic connection, checked November 11, 2026