Kortix vs Paperclip: an open-source company system or a control plane for your agents?

Canonical: https://kortix.com/blog/kortix-vs-paperclip

Paperclip is an open-source control plane that coordinates a team of AI agents you already run, and Kortix is the open-source AI Operating System that runs those agents too. Paperclip brings org charts, goals, budgets and a task board to the agents you point at it, and it pitches the result as a zero-human company. Kortix takes the same ambition one layer down: it hosts the agents, runs each session on its own cloud computer, and lands the work as a change request a person reviews. The question is whether you want a dashboard over agents something else runs, or the system that runs them.

What Paperclip gets right

Paperclip is a Node.js server and a React UI that orchestrates a team of AI agents into one organization, and it calls itself the app people use to manage agents at work (Paperclip on GitHub, checked October 2026). Its framing is exact: if OpenClaw is an employee, Paperclip is the company.

• It meets your agents where they are. Bring Claude Code, Codex, OpenClaw, Cursor, Gemini CLI, OpenCode, Hermes or any HTTP or webhook bot. The rule is that anything able to receive a heartbeat can be hired, and adapter packages extend the roster.
• It gives the team a shape. Agents get roles, titles, reporting lines and permissions, and delegation flows up and down a mixed human and agent org chart. Hiring a new agent is gated by board approval by default.
• Budgets carry hard stops. Cost is tracked per company, agent, project and goal, and an agent that reaches its limit auto-pauses with new tasks blocked (cost controls, checked October 2026).
• One dashboard for goals and history. Tasks link to company goals so agents see the why, routines run on a cron or a webhook, and every instruction, tool call and decision is written to an append-only activity log.
• It is open source and self-hosted. Paperclip is MIT-licensed, self-hosts from one npx paperclipai@latest onboard --yes with an embedded Postgres or deploys with Docker, and needs no Paperclip account (the Paperclip repository, checked October 2026).

The project moves fast. Its most recent release, v2026.1012.0, shipped on 2026-10-12 (releases, checked October 2026), and the repository carries more than 100,000 stars. Paperclip is a serious, well-built control plane, and for the job it names, organizing agents someone else runs, it does that job well.

Where an orchestration layer stops

Paperclip coordinates agents; it does not run them. The runtime, the model and the machine stay the agent's own, which is fine until the question becomes where the work executes and where the company's definition lives.

• The agents are yours to supply. Paperclip states it is not an agent framework: it does not tell you how to build agents, and the model, harness and machine remain your responsibility. Sandboxing is optional and adapter-dependent: Paperclip's own repository says sandbox providers extend execution beyond the local host, with availability set by the configured environment and adapter (the Paperclip repository, checked October 2026).
• The organization lives in Paperclip's own store. Roles, goals, budgets and skills are state inside that database. Company portability exports and imports them as packages rather than treating a git repository as the live source of truth you diff and roll back (company portability, checked October 2026).
• Work lands as a ticket in the dashboard. Paperclip's unit of work is the task with an owner, a status and a thread, and its review stages live inside the product. That governs the work there; it is not a single gate where a session's output is read as a change against main before it lands.
• There is no isolated computer per session by default. Each Paperclip run wakes an adapter through a heartbeat, the execution workspace is optional, and the host you point it at is the machine you manage.

None of this is a flaw in what Paperclip is; a control plane hands execution to somebody by design.

The company is a repo you own

Kortix makes the repository the company itself: your agents, their skills, your company memory, connector configuration and triggers are files in one git repo you own (Kortix docs). Where Paperclip keeps the organization in its database and exports it on demand, Kortix keeps it as text you can grep, diff and roll back.

A short kortix.yaml at the repo root declares how a session boots:

# kortix.yaml (excerpt)
default_agent: ops-morning
sandbox:
  image: Dockerfile
agents:
  ops-morning: { connectors: [slack, stripe] }
triggers:
  - { type: cron, agent: ops-morning, cron: '0 7 * * 1-5' }

• Grep the whole company. Agents and skills are Markdown, memory is files, and connector rules and triggers are config. Version every change, read it as a diff, roll any part of it back.
• One place to start and steer. Sessions start from the web app, Slack, Microsoft Teams, the CLI or the API, or from cron and webhooks with nobody present. Email is experimental, per project.
• The tools are in the repo too. Kortix connects 3,000+ apps in a click, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP. Connector credentials are brokered server-side and never enter the machine, and you set allow, ask or block per tool call.

Every session gets its own computer and a review gate

The part Paperclip leaves to you, Kortix runs as the platform. Each session boots its own isolated cloud computer on a branch named after the session, and the agent commits there. Thousands run in parallel on one config with no crossover between them.

That isolation sets up the unit of work. A session's changes reach main only through a change request, and merge is default-deny for agents, so a person reads the diff before it lands. Paperclip's approvals govern a task inside its dashboard; Kortix's gate is a reviewable change against the company's own repository.

• Any model, your keys. Pick a model per agent, per session or per message: any model provider with your own API keys, or the ChatGPT plan you already pay for.
• A fleet of one machine each. Every run starts from the same image with your repo and tools already on it, so there is nothing to install and no laptop to keep awake.
• Deploy where you need it. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. Free is $0 with 200 credits per month and 1 project; the team plan is $40 per seat per month with 2,500 credits per month per seat, pooled (Kortix pricing, checked October 2026).

Kortix is open source: read it, fork it, audit it, and self-host for free. The source is at Kortix on GitHub.

Where Paperclip is the better pick

If your team already runs agents such as Claude Code or Codex across several vendors and wants to organize them, Paperclip is the stronger fit. Its org chart, per-agent budgets with automatic pauses, board approvals and append-only activity log are built for exactly that job, and they are more developed than what Kortix ships for coordinating agents you run elsewhere. A team whose only problem is which of twenty agent tabs is doing what, and what it is costing, should start with Paperclip.

Side by side

When to pick which

Paperclip is the better start for a team that wants to govern a fleet of agents it already owns. Kortix is the one to run when the company's definition should live in a repository you own and every session's work should reach main through a change request. Both are open source, so the choice comes down to how much of the stack you want the platform itself to carry.

The wider comparison is in Kortix vs OpenHands. To choose a platform rather than a control plane, read how to pick an open-source AI agent platform.

Own the whole agent workforce, not just the dashboard over it. Kortix is free to start and free to self-host. Get started with open-source Kortix.