Paperclip
Paperclip is an open-source, self-hosted platform that organizes AI agents into a company: org chart, roles, goals, budgets and governance. Model-agnostic across Claude, Codex, Gemini and more, it puts you on the board of directors.
What is Paperclip?
Paperclip describes itself as the app people use to manage AI agents for work. Its premise is blunt: AI agents are getting good enough to do real jobs, but a pile of chat sessions is not a workforce. Instead of steering agents one isolated session at a time, you structure them as a company, with an org chart, roles, goals, budgets, governance and accountability. The path shown on the site has three steps: define the goal, hire the team, then approve and launch. Six capabilities carry the product, namely Bring Your Own Agent, Org Chart, Goal Alignment, Cost Control, Ticket System and Governance. Around them, four product areas address distinct audiences: an Agentic Task Manager for everyone, an Org Chart for Agents for managers, Agent Employee Training for enablers, and an Agentic OS for IT and platform teams. Paperclip does not supply the models. It is runtime-agnostic, shipping seven adapters and naming eight runtimes, from Claude, Codex, Gemini and Cursor to Hermes, OpenClaw, Pi and OpenCode, under a single admission criterion: if it can receive a heartbeat, it's hired. Agents wake on a schedule rather than running non-stop; the site illustrates a copywriter every four hours, an SEO analyst every eight and a social manager every twelve. Work stays anchored by an alignment chain running from Mission to Project Goal to Agent Goal to Task. Cost and control are first-class. Each agent carries a monthly budget cap, with a warning at 80% and an automatic stop at 100%; the sample org chart shown on the site totals 240 USD of such agent budgets. Every instruction, reply, tool call and decision is written to an append-only log. You operate as the board of directors: agents cannot hire other agents without your approval, and any agent can be paused, reassigned or terminated. The software is MIT-licensed and self-hosted, a single instance can run dozens of companies with complete data isolation, and Paperclip Labs, Inc. shipped 24 releases between 15 January and 22 July 2026.
What it does
- Set a company goal and let the CEO agent propose a strategy
- Hire agents by role: CEO, CTO, engineers, designers, marketers
- Assign tickets with an owner, a status and a discussion thread
- Cap the monthly budget of every agent
- Approve or reject hires, strategies and publications
- Run agents on scheduled heartbeats
- Trace every tool call in an append-only audit log
When to use Paperclip / When not to
A quick filter to help you decide if Paperclip is the right fit.
When to use Paperclip
- Founders and small teams running several coding agents in parallel
- Technical teams comfortable with git clone, Docker or Node.js
- Organizations that require data and API keys to stay on their own infrastructure
- Cross-functional deployments across engineering, support, sales and marketing, operations and finance
- Platform teams that need traceability, an audit log and per-agent cost caps
When not to use Paperclip
- Buyers looking for a ready-to-use SaaS, since hosted access is waitlist-only
- Non-technical users who expect a one-click sign-up, as installation requires Node.js 20 or later, pnpm and an embedded PostgreSQL
- Mobile-first teams, as there is no iOS or Android application
- Procurement-driven organizations, with no published certification (SOC 2, ISO 27001, HIPAA), no DPA, no subprocessor list and no GDPR mention
- Teams expecting models to be included, or wary of very young software: you bring your own provider keys and budgets, and the first public release dates from 15 January 2026
How to use Paperclip
A typical end-to-end flow, from setup to results.
- Install Node.js 20 or later, then enable pnpm through corepack
- Run the single onboarding command, npx paperclipai onboard --yes, and never with sudo or in a root shell, as the embedded PostgreSQL refuses to run as an administrator user
- Let the command download Paperclip, create the ~/.paperclip/ directory, initialize the embedded PostgreSQL database and start the server, which the solutions page times at around five minutes
- Open the interface at http://localhost:3100 as documented, keeping in mind that the home page illustration shows port 3000
- As an alternative shown on the open-source page, clone the repository and run docker compose up
- For a server deployment on AWS, GCP, DigitalOcean or Hetzner, point a domain at the machine and terminate TLS with Nginx, issuing and renewing the certificate with Certbot
- Provide your model provider keys as environment variables, ANTHROPIC_API_KEY and OPENAI_API_KEY
- Create a company, then hire your first agent and give it a role
- Set that agent's monthly budget, then assign it a first task
- Point your agents at the bundled SKILL.md so they can discover the context and help with the deployment
Pros & Cons
Pros
- MIT license and self-hosting: the code can be audited, and data and keys stay with you
- No vendor lock-in, since an agent's runtime can be swapped without rewriting the role
- No usage meter phoning home: you pay your model providers directly
- Hard per-agent budget caps with an automatic stop at 100%
- Strong traceability through an append-only log, tool-call tracing and skill citation
- Governance by default, where autonomy is granted rather than assumed, and cross-functional coverage documented for five business functions
- Sustained release cadence, with 24 releases in a little over six months, and substantial documentation (27 pages of API reference, 31 for the CLI, 20 for adapters, 9 for deployment)
Cons
- No published pricing and no pricing page, with hosted access limited to a waitlist
- Command-line installation is mandatory, there is no access without a terminal, and no mobile application exists
- Very young product: the domain was registered on 2 March 2026 and the first Wayback capture dates from 6 March 2026
- No postal address is published, including in the legal pages
- No published security certification, no DPA, no subprocessor list and no mention of the GDPR
- The content license in section 7 of the terms is broad and survives the end of your use, and enabling detailed telemetry explicitly authorizes model training
- Documentation and interface are English-only as far as is documented, the 74k+ GitHub stars are a claim by the site rather than a figure verified here, and the home page (port 3000) contradicts the documentation (port 3100)
Pricing & Plans
The software is free of charge: Paperclip is published under the MIT license, is self-hosted, and requires no Paperclip account. The site has no pricing page and displays no price for the product itself, and there is no usage meter, the publisher stating that you pay your model provider directly. The real cost to the user is therefore that of their own model providers and their own infrastructure. A hosted offering is mentioned through a waitlist on the home page, without a price and without a date. Readers should note that the per-agent amounts visible in the interface screenshot on the site (60, 40, 50 and 30 USD, totalling 240 USD) are example agent budgets, not prices charged for the tool.
- No commercial plan is published
- free self-hosting under the MIT license
- Hosted access is announced through a waitlist
- with no plan name
- no price and no date
Data, GDPR & hosting
A consolidated view of how Paperclip handles your data.
GDPR overview
Paperclip makes no GDPR claim anywhere: the regulation is not mentioned on any page reviewed. No EU representative under Article 27 and no data protection officer are named, and neither standard contractual clauses nor an adequacy decision are referenced. What the privacy policy in force, dated 23 July 2026, does provide: section 8 lists rights of access, rectification, erasure, objection, restriction, portability and withdrawal of consent, but qualifies them as depending on your jurisdiction; section 9 states that data may be transferred to and processed in countries other than your country of residence, including the United States, on the basis of your consent. Requests go to privacy@paperclip.ing, or failing that legal@paperclip.ing. If you self-host, you become the controller for the data your own deployment processes.
Who owns the data?
Section 7 of the terms states that you retain ownership of the content you submit. In exchange, Paperclip Labs, Inc. receives a worldwide, non-exclusive, royalty-free, transferable and sublicensable license to use that content to operate, improve and develop its Services. That license survives the end of your use for content already processed or incorporated into aggregated or anonymized datasets. The publisher keeps all rights to the Services themselves, their technology, algorithms, models and aggregated learnings. Where you self-host the open-source software, the privacy policy is explicit that your own deployment processes your data on your infrastructure, the policy covering only the Services hosted and operated by the publisher.
Reuse rights
Section 3 of the privacy policy limits collected data to operating and improving the Services, delivering requested features, usage analysis, product development, communication, security and legal obligations. Data from connected third-party accounts is not sold, not used for advertising and never used to train generalized AI/ML models; it is not read by humans except with your consent, for security or under a legal obligation, and section 4.1 commits to the Google API Services User Data Policy, including its Limited Use requirements. Section 5 states that personal data is not sold, and that sharing is limited to service providers, authorities or a corporate transfer. Telemetry follows a different rule: under section 8 of the terms, anonymized telemetry is collected automatically, while detailed telemetry requires explicit consent, and enabling it grants the publisher the right to collect, process and use that data for any purpose related to the Services, including the training and improvement of machine learning models. On a self-hosted deployment, the content you and your agents produce stays yours to reuse without asking anyone.
Data retention & training
Hosting summary
Where data lives depends on the deployment mode. If you self-host, which is the documented route, everything stays with you: the open-source page states that 100% of your data stays on your infrastructure and that your tasks, threads, documents and keys never leave it. A local install keeps them in an embedded PostgreSQL database and local files under the ~/.paperclip/ directory, while a server install places them on the cloud of your choice, AWS, GCP, DigitalOcean and Hetzner being the examples given. For the Services hosted and operated by Paperclip Labs, Inc., section 9 of the privacy policy states that data may be transferred to and processed in countries other than your country of residence, including the United States. No specific hosting country or region is named for those operated Services, and infrastructure providers are referred to generically in section 5 without a nominative list. A network measurement of the marketing domain resolves to 216.150.16.65, AS16509 Amazon.com, anycast, which describes the website rather than where customer data is hosted.
Where Paperclip works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Paperclip.
- The content license in section 7 of the terms is very broad and survives the end of your use for anything already aggregated or anonymized
- Detailed telemetry, once enabled, explicitly authorizes the collection, analysis and use of that data to train and improve machine learning models
- Liability is capped by section 10 of the terms at the greater of the amounts paid over the preceding twelve months or 100 USD, which on free software means 100 USD in practice
- Section 14 imposes Delaware law, mandatory individual arbitration and a class action waiver
- Section 13 allows suspension or termination at any time, with or without cause and with or without notice
- Agents take real actions in third-party systems connected through OAuth, so the breadth of the scopes you grant should be assessed before connecting anything
- No postal address is published anywhere on the site, there is no certification, no DPA, no subprocessor list and no GDPR mention, on a product whose domain was only registered on 2 March 2026
Setup & Integrations
Technical difficulty
Moderate for the local route, demanding for a server. Installation is command-line only: Node.js 20 or later, pnpm through corepack, then the single command npx paperclipai onboard --yes, timed at around five minutes, with documentation that explicitly reassures readers unfamiliar with a terminal. Do not run it with sudo or in a root shell, as the embedded PostgreSQL refuses to start as an administrator user, and expect a documented pitfall on Apple Silicon macOS with missing symlinks for libzstd and liblz4. The server route adds DNS, Nginx, Certbot, a dedicated user and a systemd service.
Deployment
Integrations
Behind Paperclip
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Paperclip.
Frequently asked questions
How is Paperclip different from agents such as OpenClaw or Claude Code?
Can I use the agents I already run?
What happens when an agent reaches its budget?
How do I stop an agent doing something I do not want?
Do agents run continuously?
Why not simply connect my agent to Asana or Trello?
What does a typical installation look like?
Can I run several companies on one instance?
Is it really open source, and what does it cost?
Is my data used to train AI models?
Should you pick Paperclip?
Paperclip's positioning is unusually clear and unusually well documented: it is the management layer above AI agents, not one more agent. Where most tools compete to be the smartest worker, Paperclip assumes you already run several and asks the harder question of how to run them together, through an org chart, a mission-to-task alignment chain, hard monthly budgets with an automatic stop, approval gates on hiring and strategy, and an append-only log of every instruction and tool call. The MIT license and self-hosting complete the picture, since the code can be audited and tasks, threads, documents and keys stay on your own infrastructure. The reservations are just as clear. The product is less than six months old, the domain having been registered on 2 March 2026, even if 24 releases in that time suggest genuine momentum. There is no pricing page and no named plan: the software is free, and hosted access sits behind a waitlist with neither price nor date. Nothing addresses procurement either, with no postal address published anywhere, no security certification, no DPA, no subprocessor list, and no mention of the GDPR. Two clauses deserve a legal read before adoption, namely the very broad content license of section 7 of the terms, which survives the end of your use, and the fact that turning on detailed telemetry explicitly authorizes the training of machine learning models. The natural audience follows from all this: technical teams and founders who already run several agents in parallel, are comfortable with a terminal, and want data and API keys to remain their own. For them, Paperclip offers governance and traceability that a pile of chat sessions cannot provide. For everyone else, the absence of a ready-to-use hosted service and of any compliance commitment remains the decisive obstacle.
- Choosing a selection results in a full page refresh.
- Opens in a new window.