The open-source self-hosted AI agent platform you run yourself

What separates self-hosted from managed

A self-hosted AI agent platform runs agent sessions, company memory and configuration on machines you control, so nothing about how your company works sits in a vendor’s account. Kortix, the open-source example the others are measured against, runs that way. Five properties separate a self-hosted platform from a managed one: runtime, data, configuration, model keys, and the review gate. Buyers usually name the first four and miss the fifth.

The runtime is the part you hold. Kortix self-hosts as one Docker Compose stack built from the same images its managed cloud runs, and that stack runs on a laptop, a VPS, a VPC or your own on-prem network. Agent sessions are a deliberate exception: they run on a separate sandbox provider, Daytona by default, with Platinum and E2B supported as well, so the host itself stays small.

The data and the configuration sit on your own disk. A Kortix instance keeps its Postgres database, file storage and secrets in one instance directory you can copy, and the docs count those three as the whole backup. Configuration is more than a settings file: your agents, their skills, your company memory and every connector live in one git repo you own, alongside kortix.yaml. A managed platform keeps the same material inside its database, so leaving it means losing it.

Model keys are the fourth. A self-hosted instance connects the providers you already pay for: "Any model provider with your own keys. Or the ChatGPT plan you already pay for."

The review gate is the fifth, and the one most teams skip. A managed platform draws its own boundary around what an agent may do. On a self-hosted platform you place the gate yourself, and the strongest form is a change request. On Kortix the gate is structural: "Session work reaches main through a change request. Merge is default-deny for agents." An agent works on its own branch, and a person reads the diff before it lands.

What you own, and what you still rent

Ownership is a spectrum. Self-hosting moves the runtime, the data and the configuration onto your own hardware, and two bills stay rented whatever you choose.

Hardware you do not physically own is still rented compute. A VPS is your machine in law and someone else’s hardware in fact, and buying a box ends that arrangement. Model inference is the second. Unless you run local models, every project on this page calls a hosted provider, so the token bill keeps arriving. Neither is a hidden cost, and the point is that a self-hosted platform leaves both decisions to you instead of bundling them.

The comparison table on this page measures that split across Kortix, OpenWork and Eigent. For the wider field, see open-source AI agent platforms.

The projects that self-host today

Three projects clear the bar for a real self-hosted agent platform, and each takes a different shape. The open-source Claude Cowork hub collects the rest of this site.

Kortix is an open-source AI Operating System built on one idea: your company is a git repo. Agents, skills, memory, connector configuration and triggers are files in that repo, and kortix.yaml declares what each session boots and what it may touch. One isolated sandbox per session, each with "its own isolated machine and branch," and session work lands through a change request a person approves. Any model runs on your own keys, and connectors reach "3,000+ apps in a click, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP." The code is public at Kortix on GitHub, and self-hosting is three commands (Read the docs):

curl -fsSL https://kortix.com/install | bashkortix self-host init --domain kortix.example.comkortix self-host start

On a bare Linux box a single bootstrap script installs Docker, installs the CLI and starts the stack together; on another operating system, run those three commands. The instance lands under ~/.config/kortix/self-host/<instance>/, and its updater runs once a day unless you pin a version.

OpenWork is a free, open-source desktop app for macOS, Windows and Linux, built on the OpenCode harness. In desktop mode your files stay on your machine and prompts go straight to the provider you choose, with 50+ providers and local models on your own keys. Teams can self-host the control plane or buy a private managed instance, and its own example pauses the agent before it posts a message to Slack (openworklabs.com).

Eigent is an open-source multi-agent desktop app that organizes work as spaces, sessions and tasks. Its self-hosting docs lay out three arrangements: a managed application, self-hosted with your own provider APIs, or fully local on runtimes such as Ollama, vLLM or LM Studio. You follow each agent’s progress, inspect its sources and file changes, then approve the result or ask for revisions. For Eigent against other tools, Read the comparison.

Who self-hosts, and why

Predictable cost

Per-seat pricing punishes adoption, because every person you add is another line item. Self-hosting moves the bill to what the agents do. You pay your model provider per token at published rates, and the platform costs the hardware it runs on. For Kortix, "Self-host is free," so the only new line is the box. Most self-hosted AI workspaces follow the same shape: a trigger, an isolated run, and a reviewed output.

Data you can hand to an auditor

Sessions, memory and connector configuration stay on your disks, which is the difference between showing an auditor your own logs and asking a vendor for theirs. "Secrets are encrypted at rest with a key per project," and "every action is recorded" on the instance you control.

Governed agent fleets

A security team asks two questions: what may an agent touch, and who reviewed what it changed. Kortix answers both in files. It ships "Per-resource permissions for people and agents. Roles, groups, and an audit trail," and each tool is set to allow, ask or block, down to the arguments of a single call. "An Ask holds the call until a person approves it, then the agent resumes," and the shipped default is plain: "Approval gates you set. Off until you set them." On identity, Kortix ships "SAML 2.0 single sign-on and SCIM 2.0," and "Connector credentials are brokered server-side and never enter the machine." "Thousands of agents in parallel on one config, each on its own cloud computer," and session work reaches the default branch only through a merged change request.

A checklist before you commit

Five questions settle it:

1. Does it run on my hardware? A Compose stack needs a Linux host with Docker; a desktop app needs your laptop.
2. Whose model keys pay? Your keys mean published rates; credits and seats mean a metered subscription.
3. Where does memory live? Files you can copy and back up beat a database you reach through a vendor login.
4. Is there a human gate? A change request a person merges is the strongest form; a chat log is the weakest.
5. Can I read the configuration? If agents, triggers and permissions are files, every answer is checkable.