Latest release v2026.1023.0 is live (October 23, 2026).
See what's newPaperclip is MIT-licensed and self-hosted. The control plane — org chart, tasks, approvals, budgets, and the record of every run — runs on your machine or your server. Your agents call the model providers you choose, on your keys.
$ node --version
v24.11.0
$ npx paperclipai onboard --yes
✓ Paperclip running at http://localhost:3100
→ Hire your first agent to get started
Run npx paperclipai onboard --yes on Node.js 24.11 or later. The default install is a single Node process with an embedded PostgreSQL database and local file storage under ~/.paperclip. It starts in local_trusted mode, bound to 127.0.0.1, and serves the app at http://localhost:3100. No sales call and no waitlist — the install guide is the onboarding.
Self-hosting stores the company record in your database. Paperclip doesn’t run the model, though. Each agent’s runtime, such as Claude Code or Codex, sends the task context, the files it reads, and its tool results to the model provider you configure. Connectors call the services they connect to, and anonymous usage telemetry is on until you turn it off.
Publish “Comparing agent runtimes” to the blog
Owning the stack means owning its upkeep. You supply model access — a provider API key or a runtime’s own subscription login — and pay the provider directly. You back up the database, uploaded files, and the secrets master key; lose that key and you lose the secrets it encrypts. Before anyone else connects, switch to authenticated mode, and put HTTPS in front before you expose it to the internet. You upgrade when you choose.
Paperclip provides AI agent management through roles, assigned tasks, budgets and configured review stages. It coordinates the runtimes you choose; it does not replace the framework used to build an agent. When choosing an AI agent platform for local AI agents, check both where the management record lives and where model calls go.
The tools below address different layers. Setup burden describes the work their documentation asks you to do; it is not a measured installation time or score. Commercial services and enterprise editions have separate terms and feature sets.
| Product | Deployment model | Best fit | Multi-agent support | Governance depth | Setup burden | Limitations |
|---|---|---|---|---|---|---|
| Paperclip | Self-hosted control plane; agent runtimes configured separately. | Managing ongoing work across agents and runtimes. | Roles, reporting lines and linked delegated tasks. | Configured task reviews, scoped run budgets and separate connector policies. | Install the server, connect runtimes and maintain storage, access and backups. | Recorded-cost budgets can overshoot; task review does not intercept every external call. |
| LangGraph | Library in your application; deployment services are a separate layer. | Developers building stateful agent workflows. | Graphs can coordinate workers and parallel steps. | Checkpointed interrupts support human input where developers place them. | Write graph logic and configure persistence and deployment. | Application logic and operating policy remain your responsibility; paid deployment features are separate. |
| CrewAI | Python framework; Cloud is a separate offering. | Role-based crews and event-driven flows in an application. | Sequential crews or a configured hierarchical manager. | Task human input and manager validation are available when configured. | Define agents, tasks, tools, models and process configuration. | A crew’s manager is an application mechanism; hosted platform features do not automatically belong to the MIT core. |
| OpenHands | Current Agent Canvas can connect local, remote or cloud backends; commercial Cloud/Enterprise are separate. | Coding-agent conversations and development automations. | Multiple agent backends and automations; coordination depends on the selected components. | Self-hosted Canvas documents API-key access and network hardening; commercial controls vary by offering. | Configure Canvas, its backend and host access; harden a reachable server. | Agents can access the backend host’s files, shell and network. Check the current component rather than older Local GUI instructions. |
| Dify | Docker-based self-hosting or managed Dify Cloud. | Visual AI applications, chatbots and workflows. | Agent nodes compose within a single workflow. | Workspace roles, configured tool credentials, iteration limits and Human Input nodes that pause a workflow for review. | Configure containers, providers, data sources and application flows. | Modified Apache terms add multi-tenant and frontend branding conditions; check intended use. |
| Flowise | Self-host the published code; historical Cloud instructions need current availability checks. | Maintaining or adapting visual AI workflows. | Agentflow V2 supports supervisor/worker patterns and shared state. | Configured human-input checkpoints and tool approval; enterprise code has separate terms. | Configure the server and flows, then take responsibility for ongoing maintenance. | The repository is archived; maintainers announced end of life. Do not assume ongoing core-team support or hosted availability. |
| n8n | Self-hosted editions or managed n8n Cloud. | AI steps within business and service automations. | AI Agent Tool nodes let a root agent delegate to specialist agents. | Configured human review can gate selected agent tools; admin features vary by edition. | Configure hosting, credentials, workflow nodes and review channels. | Covered code uses the Sustainable Use License; .ee files require enterprise licensing. Check permitted use and edition. |
.ee files require enterprise licensing. Check permitted use and edition.License boundaries. If you are comparing open source AI agents, check the actual component license. Paperclip, LangGraph and CrewAI have MIT-licensed cores. OpenHands Agent Canvas is MIT-licensed; its component repositories and commercial offerings need separate checks. Dify uses modified Apache-based terms with additional conditions. Flowise separates Apache 2.0 code from named commercial directories and files. n8n uses its Sustainable Use License for covered code and separate enterprise terms for .ee files. Dify’s and n8n’s terms make them source-available with conditions rather than OSI-approved open source; Flowise is Apache 2.0 outside its commercial files.
For Flowise’s maintenance status, see the maintainers’ end-of-life announcement.
Choose a framework when your main job is building one application’s execution logic. Use a single coding agent when one operator can manage the task and review its result directly. Add a management layer when work needs retained ownership, handoffs and review across a continuing team. These choices can coexist.
Yes. Install locally or on a server. Use authenticated mode beyond a trusted local machine and HTTPS for public access. You maintain storage, backups and upgrades. See installation and deployment modes.
No. The company record stays in your configured storage, but model calls send runtime context to your chosen provider and connector calls reach connected services. Anonymous usage telemetry is enabled by default and can be disabled. Review the data path above.
Adapters connect runtimes such as Claude Code and Codex. Process, HTTP and external adapters provide other integration paths. Each needs its own compatible runtime and credentials. See the adapter overview.
Only where configured. Task review and approval stages govern completion. Connector actions have separate Allowed, Ask first and Off settings. Normal new active actions start as Allowed, including writes. Set Ask first for human approval or Off to prevent use. See action permissions.
Yes. Company, agent and project policies can stop new runs when recorded spend reaches a cap. Accounting delay and in-flight work can cause overshoot; warn-only policies keep spending. See budgets and costs.
Inspect linked tasks, run records, reported costs and recorded review decisions. The detail available depends on what the adapter captures; this is not a record of every action outside Paperclip. See execution policy and adapters.
The project is MIT-licensed and has crossed 100k stars on GitHub — adapters, skills, and plugins come from the community as well as the core team. File issues, read the source when the docs fall short, and ship the integration you need instead of waiting for a roadmap.
“The honest answer to “where does our data go” is a list, not a slogan: the company record is stored on your server, and whatever an agent reads goes with its model calls to the provider you chose. Open source is how you check that list yourself.”
One command to a running company, and the source to check what it does.