ARTICLE / SEPTEMBER 18, 2026
Open Source Claude Cowork Alternative: Self-Hosted Guide
Compare Kortix, Claude Cowork, and five self-hostable alternatives by license, self-hosting, model choice, configuration ownership, and review gates.
By Kortix Blog · Published: September 18, 2026
An open source Claude Cowork alternative is software you can run and inspect yourself instead of subscribing to a hosted assistant. For a team, the decision usually turns on three separate questions: what the license lets you do, who operates the software, and where the work and your data actually run. This guide answers those questions for Kortix, Anthropic's Claude Cowork, and five other projects cited as self-hostable Cowork replacements, and it names the cases where Kortix is the wrong choice. Open Source Definition
Documentation reviewed: September 18, 2026. This is a documentation-based comparison of licenses and published product documentation, not a hands-on benchmark, security certification, or legal opinion. Vendor and product copy is labeled as such.
TL;DR: the short verdict
Kortix is the strongest documented fit when a team wants to self-host an AI coworker, keep configuration in a git repository it owns, and require human review before work lands. Kortix's repository carries Elastic License 2.0, so this guide classifies it as source-available rather than OSI open source, even though the project's own copy calls itself open source. Claude Cowork is Anthropic's hosted product with no documented self-host option and no published source license. The five other projects in this shortlist are desktop or local-first tools under permissive licenses, with the exception of OpenWork's source-available enterprise directory.
License, self-hosting, and model choice
| Option | License classification | Self-host | Model choice | Evidence (reviewed 2026-09-18) |
|---|---|---|---|---|
| Claude Cowork (Anthropic) | Proprietary hosted product; no published source license | Not documented; sessions run in the cloud on Anthropic's servers (beta) | Claude models via a Claude account, or Amazon Bedrock, Google Cloud, or Microsoft Foundry | Product page · Help Center |
| Kortix | Elastic License 2.0 — source-available, not OSI open source | Yes — runs the whole control plane on your infrastructure from one Docker Compose stack; a sandbox provider is still required and air-gapped topologies are scoped separately | Any provider with your own API keys through the in-stack gateway | Repository · LICENSE · Self-hosting |
| Eigent | Apache License 2.0 — OSI open source | Yes — local deployment with a local backend and local model integration | Model agnostic: cloud APIs, enterprise gateways, or local inference (vLLM, Ollama, LM Studio) | Repository · LICENSE |
| OpenWork | MIT outside ee/; OpenWork EE License (source-available) inside ee/ | Yes — desktop app; MIT core, with subscription terms for the Den control plane | Bring your own API keys | Repository · LICENSE |
| Open Cowork | MIT — OSI open source | Yes — local desktop app for Windows and macOS with VM-level sandbox | Claude, OpenAI-compatible APIs, and Chinese models with your own keys | Repository · LICENSE |
| Open Claude Cowork | MIT — OSI open source | Runs locally as a desktop app; requires your own Anthropic and Composio API keys | Claude Agent SDK or OpenCode | Repository · LICENSE |
| OpenClaw | MIT — OSI open source | Yes — local Gateway on your own hardware | Hosted and local providers; models are swappable plugins | Repository · LICENSE |
Configuration ownership, review, and best fit
The "best fit" column is this guide's editorial assessment, not a vendor ranking or a benchmark result.
| Option | Configuration ownership | Review / approval gate | Best-fit assessment |
|---|---|---|---|
| Claude Cowork | Inside the Claude product: folders, tools, permissions, and organization settings | Plan preview and approval prompts; enterprise permission controls and OpenTelemetry activity streams | Teams that want a managed coworker inside the Claude ecosystem and do not need to self-host |
| Kortix | Files in a git repository you own | Change requests for repository work; configurable action approval gates; merge deny-by-default for an agent | Teams that want to own the control plane, the configuration, and the review path |
| Eigent | Local application configuration | Task monitoring; the repository lists SSO and access control among enterprise features | Teams that want a multi-agent desktop workforce with local-first execution |
| OpenWork | Local workspaces and skills; Den is the organization control plane | Not established in the reviewed repository README | Teams that want a desktop app plus shareable skills and MCP connections across existing agents |
| Open Cowork | Local application configuration | Workspace confinement and VM isolation; a per-action review gate is not established in the reviewed README | Individual users who want one-click desktop automation inside a VM sandbox |
| Open Claude Cowork | .env and SKILL.md files | Not established in the reviewed repository README | Teams already using Composio that want a local Claude Agent SDK desktop client |
| OpenClaw | Local configuration and workspace files | Sender pairing approvals; sandboxing is optional and documented as a setup step, not a universal pre-action gate | Users who want a chat-driven local assistant across messaging channels |
What Claude Cowork is
Claude Cowork is Anthropic's hosted agentic product that brings Claude Code's agentic architecture to knowledge work without a terminal. Anthropic describes Cowork as taking a goal, working across chosen files and tools, and returning polished work for review, and its help center documents direct local file access on desktop, sub-agent coordination, long-running tasks, scheduled tasks, projects, and browser actions. As of September 18, 2026, Anthropic lists Cowork on paid plans (Pro, Max, Team, Enterprise) across desktop, web, mobile, and the Chrome side panel, with availability varying by surface. Product page · Help Center
Cowork is designed around handing off a task rather than asking one question at a time. The product page describes a handoff from the message box next to Chat: Cowork works in the folders and tools you choose, shows each step it takes, runs the same session across web, desktop, and mobile, keeps working when you close your laptop, schedules recurring reports, and can split a large project into chunks that run together. A built-in browser opens sites, reads pages, and fills forms inside a Cowork side panel rather than your own browser, and plugins bundle tools, knowledge, and workflows with skills, connectors, and sub-agents as the extension points. Product page
Administrators get their own controls. The product page says Teams and Enterprise admins can set access by team, manage spend limits, and control tool permissions by department, with activity streams to a SIEM through OpenTelemetry, and the help center says Cowork activity on web and mobile is captured in the Compliance API. Anthropic also documents safety limits for the agentic design, including network egress permissions that do not apply to the web fetch, web search, or MCP tools. Product page · Help Center
Claude Cowork runs tasks in the cloud. Anthropic's help center says Cowork runs on Anthropic's servers in an isolated environment (in beta), with sessions and files saved to the Claude account, and the product page lists a Claude account or Amazon Bedrock, Google Cloud, or Microsoft Foundry as where data lives. Neither official page documents installing or operating Cowork on infrastructure you control, and Anthropic publishes no source license for it, so Claude Cowork is not an open-source or self-hostable option. Help Center · Product page
How this article classifies open source, source-available, and self-hosted
The three labels answer different questions. Open source is a statement about license rights; the Open Source Initiative's definition requires free redistribution and no discrimination against fields of endeavor, among other criteria. Source-available means the code is public but the license imposes restrictions the Open Source Definition does not allow. Self-hosted says who operates the software, regardless of license. The companion open source AI agent platform guide covers the general framework; this section applies it to the Cowork alternatives compared here. Open Source Definition
Applied to this shortlist: Claude Cowork is hosted and proprietary; Kortix is self-hosted and source-available; Eigent, Open Cowork, Open Claude Cowork, and OpenClaw are self-hosted and OSI open source; OpenWork is self-hosted with an MIT core and a source-available enterprise directory. Elastic License 2.0
Kortix is the case most often mislabeled. Its repository describes itself as "the open-source AI Management System," but the repository's LICENSE is Elastic License 2.0, which forbids offering the software to third parties as a hosted or managed service and forbids circumventing license-key functionality. That hosted-service restriction limits a field of endeavor and therefore conflicts with the Open Source Definition, so this article classifies Kortix as source-available, not open source. Repository · Kortix LICENSE · Elastic License 2.0
What actually changes when you leave Claude Cowork
Four things change when a team moves from Cowork to a self-hosted alternative: who holds the model credentials, where the work and files live, who owns the configuration, and where the review gate sits. With Cowork, model access runs through a Claude account or a cloud provider, sessions and files are saved to the Claude account, configuration lives inside the product, and the approval prompts are product settings. A self-hosted alternative moves those four decisions onto infrastructure and files the team controls, which is also why it adds operational work. The rest of this article compares Kortix and the shortlist on exactly those four axes. Product page · Help Center
Kortix vs Claude Cowork, capability by capability
Kortix is an AI management system that keeps agents, shared skills, company memory, and connectors in one git repository, runs each session in an isolated sandbox on its own branch, and lands work through a change request a human approves. The repository describes Kortix as "the leading open-source alternative to Claude Cowork and ChatGPT Work" and as an "open-source AI Management System." Those are vendor positioning statements, not independent verification, and the LICENSE classification above governs the open-source question. Repository
Model choice and provider keys. Claude Cowork uses Claude models through a Claude account or Bedrock, Google Cloud, or Microsoft Foundry. Kortix's self-hosting page says every model call routes through a gateway running on your own box, connecting Anthropic, OpenAI, Google, Groq, xAI, DeepSeek, Mistral, Bedrock, and OpenRouter, or a ChatGPT or Copilot subscription you already hold. That Kortix description is vendor copy. Product page · Kortix self-hosting
Where it runs. Claude Cowork runs on Anthropic's servers in the cloud (beta). Kortix's self-hosting page says the whole control plane — accounts, projects, repos, secrets, connectors, policies, and audit — runs inside your network from one Docker Compose stack built from the same images as the managed cloud, with sandbox compute from Daytona by default or Platinum or E2B. The repository adds that self-host start pulls images from Docker Hub, so this is a self-hosted install rather than a disconnected one, and the self-hosting page says air-gapped topologies are scoped with Kortix rather than self-served. Help Center · Kortix self-hosting · Repository
Configuration ownership. Claude Cowork keeps configuration inside Anthropic's product, including chosen folders, tools, permissions, and organization settings. Kortix keeps configuration as files in a git repository the customer owns; the repository describes those files as versioned, diffable, and shared across the company. Product page · Repository
Isolation. Claude Cowork runs tasks in an isolated environment on Anthropic's servers. Kortix gives every session its own disposable Linux sandbox on its own branch, and the repository says isolation depends on the sandbox provider: the default provider runs containers, while the Platinum provider runs microVMs. Help Center · Repository
Human review. Claude Cowork shows each step and lets you redirect or stop, and the product page says Claude asks before acting when permissions are enabled, with enterprise admins configuring permissions. Kortix lands repository work only through a change request a human approves and documents configurable action approval gates, with merge deny-by-default for an agent. Product page · Repository
The shortlist: other self-hostable Cowork alternatives
The five projects below are commonly cited as Cowork alternatives. Each entry is limited to what the project's own repository and LICENSE state; capability claims from third-party listicles were not treated as proof, and no adoption or pricing figures are reproduced here.
Eigent: multi-agent desktop under Apache-2.0
Eigent is a desktop application that runs a multi-agent workforce locally and is built on the CAMEL-AI framework. Its repository states the project is licensed under Apache License 2.0 and documents a fully standalone local deployment with a local backend and local model integration through vLLM, Ollama, or LM Studio. It supports both a single-agent harness and coordinated multi-agent workflows, ships built-in browser and terminal toolkits, integrates MCP and skills, and offers a cloud backend as an alternative to the local one. Apache-2.0 is an OSI-approved open-source license. Best fit: teams that want a multi-agent desktop workspace with local execution and bring-your-own-model setup. Repository · LICENSE
OpenWork: desktop app and shareable skills, directory-split license
OpenWork is a desktop app for macOS, Windows, and Linux, built on OpenCode, that positions itself as an open-source alternative to Claude Cowork and Codex and exposes skills, MCP connections, and Google Workspace or Microsoft 365 capabilities through an MCP server. The repository says the desktop app is not required: you can add one OpenWork MCP server to Codex, Claude Code, Cursor, or another compatible agent and reuse the same skills and connections, while the Den control plane provisions inference, manages members and teams, sets desktop policies, and publishes skills through marketplaces. Its license is directory-split: everything outside ee/ is MIT, while ee/ (the OpenWork Den control plane) uses the source-available OpenWork EE License, which is free for organizations of up to five users, free to evaluate for 30 days, and converts each release to MIT two years after publication. Best fit: teams that want a desktop app with shareable skills across existing agents and can accept the EE boundary for organization controls. Repository · LICENSE
Open Cowork: one-click desktop clone under MIT
Open Cowork is a desktop application for Windows and macOS that wraps Claude Code, OpenAI, Gemini, and DeepSeek models in a GUI with one-click installers. Its repository documents VM-level sandbox isolation using WSL2 on Windows and Lima on macOS, a path-restricted baseline when no VM is available, a built-in skills system for PPTX, DOCX, XLSX, and PDF output, MCP connectors for browsers and Notion, GUI automation, and remote control through Feishu and Slack. It supports your own OpenRouter, Anthropic, or other API keys, and its LICENSE is MIT. Best fit: individual users who want a one-click desktop cowork app with VM sandboxing and their own keys. Repository · LICENSE
Open Claude Cowork: Claude Agent SDK plus Composio integrations under MIT
Open Claude Cowork is a desktop chat application that runs on the Claude Agent SDK or OpenCode and connects to more than 500 apps through Composio's tool router. Its repository documents persistent sessions, real-time streaming, tool visualization, skills support, browser automation, scheduling, and a companion messaging bot alongside the desktop client. Setup requires your own Anthropic API key and a Composio API key, and its LICENSE is MIT. Best fit: teams already using Composio that want a local Claude Agent SDK desktop client with broad SaaS integrations. Repository · LICENSE
OpenClaw: chat-driven local assistant under MIT
OpenClaw is a local AI assistant that runs a Gateway on your own hardware and reaches you through Discord, iMessage, Slack, Teams, Telegram, WhatsApp, and other channels. Its repository describes the Gateway as the local control plane for sessions, tools, events, and channel connections, with companion apps and device nodes adding voice, camera, and screen actions. It states that state, memory, and credentials live on your hardware, that models and agent harnesses are swappable plugins, and that the project has no paid tier or hosted service; it also warns that tools run on the host for the main session unless sandboxing is configured. Its LICENSE is MIT. Best fit: users who want a chat-driven personal or small-team assistant with local state. Repository · LICENSE
Where Kortix is not the right fit
Kortix is not the right fit for a non-technical solo user who wants zero setup. Self-hosting is documented as a CLI-driven Docker Compose deployment, and the repository says self-host start pulls images from Docker Hub, so it is not a disconnected or click-to-run product. A hosted product or a one-click desktop app is a better fit for that reader. Kortix self-hosting · Repository
Kortix is not the right fit for a team that wants a fully managed service with no operational ownership. Kortix does offer a managed cloud, with a published Free tier and a Team tier at $40 per seat per month, but if you will not operate a control plane at all, a hosted product such as Claude Cowork fits that requirement more directly. Kortix pricing · Product page
Kortix is not the right fit for an air-gapped requirement treated as self-serve. The self-hosting page says air-gapped and other fully isolated topologies are scoped with Kortix rather than self-served, and that choosing a sandbox provider is a required decision. Kortix self-hosting
Kortix is not the right fit if you need enterprise identity and audit controls without an enterprise key. The self-hosting page states that SAML SSO, SCIM directory sync, custom roles, groups, and audit read require an enterprise key. Kortix self-hosting
Kortix is not the right fit if an OSI-approved permissive license is a hard requirement. Elastic License 2.0 is source-available and restricts hosted or managed-service use, so an MIT or Apache-2.0 option such as Eigent, Open Cowork, or OpenClaw is the more honest starting point. Kortix LICENSE · Elastic License 2.0
How to evaluate a self-hosted Cowork alternative
The four checks below are a proposed method, not testing this article performed.
Define the Cowork task you are replacing
Write the task you would hand to Cowork, such as "prepare a customer briefing from approved records and propose a CRM update." Then decide which documented Cowork behaviors must survive the switch: direct file access, scheduled tasks, sub-agent coordination, and a review step before irreversible actions. An alternative that matches only the chat surface does not replace the task. Help Center
Check the license against your deployment
Read the LICENSE, not the README headline. If you plan to expose the capability to other teams as a service, Elastic License 2.0's hosted-service restriction is the clause that matters; if unrestricted redistribution is a hard requirement, prefer MIT or Apache-2.0. Use the Open Source Definition as the test. Kortix LICENSE · Open Source Definition
Check where the work runs and what it can reach
Ask what runs on your infrastructure and what still calls out. Kortix requires a sandbox provider and scopes air-gapped topologies separately, and its images are pulled from Docker Hub. Request the dependency map for model inference, execution sandboxes, connectors, and updates. Kortix self-hosting · Repository
Check the review gate before irreversible actions
Find the exact gate for the sensitive action, and whether it is a repository merge, a tool call, or an outbound message. Kortix documents change requests for repository work and configurable action approval gates; the desktop projects reviewed here do not document a universal pre-action gate. Once you have selected, use the reliable AI agent workflows guide for implementation controls rather than expanding this comparison into a reliability tutorial. Repository
Frequently asked questions
Is Claude Cowork open source?
No. Anthropic publishes Claude Cowork as a hosted product on paid plans, and its help center says Cowork runs on Anthropic's servers in the cloud (in beta). Neither the product page nor the help center documents a self-hosted deployment or a published source license, so Claude Cowork is neither open source nor self-hostable under the reviewed documentation. Product page · Help Center
Is Kortix OSI-approved open source?
No. Kortix's repository README says "open source," but the repository's LICENSE is Elastic License 2.0, which prohibits providing the software as a hosted or managed service and prohibits circumventing license-key functionality. That restriction conflicts with the Open Source Definition, so this guide classifies Kortix as source-available and self-hostable, not OSI open source. Kortix LICENSE · Open Source Definition
Can you self-host a Claude Cowork alternative?
Yes. Self-hosting describes who operates the software, and several alternatives in this shortlist run on your own machine or infrastructure. Kortix documents a self-hosted control plane via one Docker Compose stack, Eigent documents a standalone local deployment, Open Cowork and Open Claude Cowork are local desktop apps, and OpenClaw runs a local Gateway. Check each project's license and dependencies before committing. Kortix self-hosting · Eigent · OpenClaw
Does Kortix work with Claude models?
Yes. Kortix's self-hosting page says every model call routes through a gateway on your own box and lists Anthropic alongside OpenAI, Google, Groq, xAI, DeepSeek, Mistral, Bedrock, and OpenRouter, or a ChatGPT or Copilot subscription you already hold. This is Kortix's own vendor documentation, not an independently verified integration test. Kortix self-hosting
How is Kortix different from ChatGPT Work?
This article compares Claude Cowork. Kortix's repository positions itself against both Claude Cowork and ChatGPT Work, but that comparison is vendor copy and this guide did not independently verify ChatGPT Work's capabilities. If ChatGPT Work is a candidate, verify it against OpenAI's own documentation and apply the same four checks above. Repository
What is the difference between open source, source-available, and self-hosted?
Open source describes license rights that meet the Open Source Definition. Source-available describes public code whose license imposes restrictions outside that definition, such as Elastic License 2.0 or OpenWork's EE License. Self-hosted describes who operates the software, regardless of license. A product can be self-hosted and source-available at the same time, as Kortix is. Open Source Definition · Kortix LICENSE
Do I need an enterprise key to self-host Kortix?
Not for core self-hosting, but several governance features require one. Kortix's self-hosting page states that SAML SSO, SCIM directory sync, custom roles, groups, and audit read require an enterprise key, while the base stack, model gateway, and per-session sandboxes are part of the self-hosted deployment. Kortix self-hosting
Next step: read the license and the self-host documentation
Choose a workload first, then verify the license rights, deployment dependencies, and review gates of the candidate you will actually run. For Kortix, start with the current LICENSE and the official self-hosting documentation; for the broader platform-and-framework decision, read the open source AI agent platform guide, and for hosted-runtime trade-offs, read the OpenAI Agents API analysis for teams.