Open-source Kortix vs AnythingLLM: the workspace ends, the repo starts
AnythingLLM is a local-first app for chatting with your documents; Kortix is the open-source system a team owns, with shared memory and reviewed agent work.
AnythingLLM is a good answer to one question, what this folder of documents says; Kortix, the open-source CompanyOS, is the choice when the answer has to become work other people can find, run and approve. AnythingLLM runs on your own hardware and cites the page an answer came from, so a person can be chatting with their files within minutes. Each one is open source, and each runs on models you pick. The split arrives with the second person. They need the context the first person taught the app, and someone has to sign off before an agent touches a live system.
What AnythingLLM does well
AnythingLLM is Mintplex Labs' all-in-one, local-first app, and the description holds up. Documents in, cited answers out, the model running on your own hardware if you want it there (AnythingLLM on GitHub).
- Running in minutes. A desktop build for Mac, Windows or Linux, a Docker image, or a cloud template gets the app running.
- Citations in the answer. PDF, TXT and DOCX uploads pass through a collector pipeline, and a reply points back at the passage it used.
- Agents without a code step. A no-code builder puts agents inside the workspace and connects MCP servers as tools.
- Model and store choice. It works with popular closed and open-weight providers and several vector databases, and it ships a full developer API.
- MIT licensed. The project is MIT licensed, so you can read, modify and ship it (AnythingLLM LICENSE).
The workspace is the boundary
The design is consistent. A workspace holds its documents, its memories and the agents that work on them, and it keeps them together. That same boundary is where a team runs into trouble.
- Memory is personal. AnythingLLM keeps up to 20 workspace memories and five global ones per account, and in multi-user mode nobody sees another person's set (AnythingLLM memories). What one analyst teaches it stays with that analyst.
- Multi-user means Docker. Multi-user support and permissioning ship in the Docker build, and the desktop app serves one person.
- One instance, many agents. Agents run inside the workspace app, so an agent that installs a package or deletes a file works on the machine everyone else is using.
- Nowhere to review. Finished work stays in the thread that produced it. Nothing opens a change request that someone reads as a diff before an agent reaches a live system.
For one person with a folder of contracts, every item on that list is the right call. A desktop app should stay simple. The list bites only when a team starts treating the app as company memory, a job it was never built for.
Memory that outlives the person who taught it
The gap shows up as repeat work. A second analyst opens the same documents and gets the same first answers, because the corrections the first analyst made sit in that person's own account.
Kortix puts the company in one git repository you own: agents, shared skills, memory, connector configuration and triggers, all plain files (Kortix on GitHub). You can grep the whole company and roll back a bad week, because that configuration is version controlled and does not sit inside a product database.
A new teammate clones the repository and gets the whole company: agents in Markdown, runtime settings in kortix.yaml. One team's finding sits in the files for the next team to read.
A computer per session, a diff per change
One isolated sandbox per session. An agent can install packages, run code and wreck the place while everyone else keeps working, because each session has its own isolated machine and branch. Outside git, the damage stays in that session. Merge is default-deny for agents: a change an agent makes in a session arrives as a change request a person approves, unless an admin granted it the merge capability. A change made from the dashboard can commit straight to the default branch.
A session starts from that repository, so the newest hire's agent already has the memory the last hire wrote. The folder does not need re-teaching. The diff is the record a reviewer works from: which file changed, and what it said before.
Where AnythingLLM is the better pick
For document chat, AnythingLLM is the better pick, and it is not close on setup time. One desktop download gets a person chatting with their own files. Nothing needs a git remote, and the MIT license lets you change the whole app. A small team that wants private answers over a shared drive, with no review step in the middle, will get there with less machinery.
Side by side
| Dimension | Kortix | AnythingLLM |
|---|---|---|
| Open source | Yes | Yes, MIT |
| Where state lives | Files in one git repo you own | Per-user memories in the app |
| Memory across a team | Shared files every agent reads | 20 per workspace, 5 global, per account |
| Multi-user | Per-user and per-agent permissions | Docker build only |
| Agent execution | Isolated machine per session | Inside the workspace app |
| How work lands | Change request; agents merge only by admin grant | Answers stay in the thread |
| Self-host | Laptop, VPS, VPC or on-prem | Local, Docker or a cloud template |
Two questions decide most of it
Who needs what the last person taught the agent? And who reads an agent's change before it reaches something live? If the honest answer to both is nobody, AnythingLLM is the lighter tool and you should take it.
If either answer names a person, the workspace runs out of room. A company that needs shared memory, a computer per session and a review gate keeps all three in a repository it owns, and Kortix is built for that.
If you are still mapping the field, the self-hosted AI workspace roundup weighs six options for a team. The Open WebUI comparison covers the chat-first route beside this one.
Free to start, free to self-host. Get started with Kortix.