What is an open source AI agent platform?
An open-source AI agent platform is software you can read, run, modify, and self-host that gives AI agents everything they need to do real work: a harness to plan and act, an isolated machine to run on, memory that persists, connectors to your systems, permissions, and an interface to start and steer them. Kortix is the open-source AI Management System, the leading open-source alternative to Claude Cowork and ChatGPT Work, built so a company owns every layer: any model, your keys, self-hosted or managed cloud.
Model, framework, harness, platform: what each one is
An AI model is the trained network that reads an input and produces text, a plan, or a tool request. GPT, Claude, and Gemini are models. A model on its own answers a question; it does not hold a task open across steps, remember what happened last week, or touch your systems.
An AI agent framework is a code library that gives developers primitives for building agents, such as tool calls, multi-step loops, and state. LangChain, CrewAI, and AutoGen are frameworks. A framework leaves deployment, identity, isolation, and governance to the team that adopts it, so the team writes code and owns everything around the agent.
An AI agent harness is the runtime layer that wraps a model and runs the loop around it: it calls the model, routes tool requests, keeps state, injects context, applies permissions, runs actions in an environment, and records what happened. The harness is what turns a model that answers into an agent that finishes a task. In the standard formula, an agent is a model plus a harness (Databricks).
An AI agent platform is a product that bundles a harness with the operational layers around it: isolated execution, persistent memory, connectors to outside systems, permissions, and an interface for people to start and steer agents. A framework is code you assemble; a platform is a system you run. The practical difference is how much your team builds versus what arrives already working.
An open-source AI agent platform is a platform whose code and configuration you can read, run, modify, and self-host. That last property changes the ownership math: instead of configuration living in a vendor's database, it lives in files you hold.
The six layers a real platform provides
Most tools hand you one layer and rent you the rest. A platform is all six, and they read in order: the harness runs the agent, the machine isolates it, memory persists it, connectors reach outward, governance constrains it, and the interface starts and steers it.
The agent harness
The agent harness is the control loop that makes a model act: planning, tool use, and multi-step runs that finish. Kortix runs its harness on OpenCode, configured by a file in the repo, and sets allow, ask, or block per tool down to a single shell command. The harness is open source, so it is never the layer a team is locked into.
Isolated execution
Isolated execution gives every agent session its own machine, so one run cannot read or break another. Kortix boots an isolated Linux sandbox per session on its own branch; the agent can install, run, and break anything, and only what it commits survives. Session id, sandbox id, and branch name are the same string, and thousands run in parallel on one config.
Memory
Memory is what a platform keeps between sessions so agents do not restart from zero. In Kortix, memory is files in the same git repo as the agents: skills written once are shared into every session, and memory accumulates as the company learns. Memory is part of the diffable company, not a hidden vector store owned by a vendor.
Connectors
Connectors are how agents reach the systems a company already runs on. Kortix includes 3,000+ apps plus MCP, OpenAPI, Postman, GraphQL, and raw HTTP, and it brokers connector credentials server-side through one scoped token so the raw key never enters the sandbox. Each tool call is ruled allow, ask, or block.
Governance and permissions
Governance is who and what an agent may touch, and what a human gets to review. Kortix sets per-resource permissions for people and agents, encrypts secrets at rest, injects a granted secret at runtime, and lands every change as a change request a human reads as a diff before merge. Merge is deny-by-default for an agent, so work reaches main only through a reviewed change.
Interface
The interface is where people start, watch, and steer agents. Kortix starts the same session from the web app, Slack, Microsoft Teams, email, mobile, CLI, or API, and cron schedules and signed webhooks can start a session with nobody asking. Started from anywhere, the work still lands through the same change-request gate.
What "open source" means for a platform
Open source on a platform means the code and the configuration are yours to read, run, modify, and self-host rather than settings inside a vendor's database. In Kortix, the whole company is one git repo: agents and skills are markdown, memory is files, and kortix.yaml declares the machine image, the connectors, and the triggers. A team can grep the whole company, diff any change, and roll any part of it back. Self-hosting runs on a laptop, a VPS, your own VPC, or your own on-prem network, so the platform does not depend on one vendor's cloud. Kortix is open source (Elastic License 2.0) — self-host, read and modify the code.
Ownership matters because the closed alternatives are becoming company operating systems too. Claude Cowork and ChatGPT Work are heading for systems an entire company runs on, and you will never own those. Kortix exists so every agent, every skill, the memory, and the whole configuration stay files the company holds. When you want to run it yourself, the self-hosting guide covers the commands.
How to evaluate an open-source AI agent platform
Score a platform against six criteria before you commit to it. Each criterion asks whether the platform gives you ownership and control or keeps them on the vendor's side.
| Criterion | The question to score | What good looks like |
|---|---|---|
| Ownership | Are agents, memory, and config files you hold? | One git repo you can grep, diff, roll back |
| Model freedom | Can you point each agent at any model with your own keys? | Any provider or OpenAI-compatible endpoint, per agent |
| Self-hosting | Can it run on your own machine, VPC, or on-prem? | Self-host or managed cloud, same config |
| Connectors | Does it reach your existing tools with brokered credentials? | 3,000+ apps plus MCP, OpenAPI, GraphQL, HTTP |
| Governance | Can you scope permissions and review every change? | Allow, ask, or block per call; review before merge |
| Cost shape | Do you pay per seat, or for infrastructure and model use? | Self-host on your own infrastructure; cloud optional |
A tool that scores on one or two criteria is a component, not a platform. The strongest candidates score on all six, and the ownership row matters most: if the config is not in files you hold, none of the other five is fully yours. For a side-by-side view of the field, see the best open-source AI agent platforms.
Where Kortix fits
Kortix is the recommended platform for a team that wants to own the whole stack, and it is built to score on all six criteria without assembling them from separate vendors. Kortix is one git repo you own, each agent can run any model with your own keys, and it self-hosts on a laptop, a VPC, or on-prem or runs as managed cloud. It ships 3,000+ connectors with credentials brokered server-side, scopes every tool call with allow, ask, or block permissions behind a change-request gate, and lets self-hosting run on your own infrastructure rather than a per-seat tax.
Get started with open-source Kortix at Kortix, or read the FAQ for the questions teams ask before they choose a platform.