The question here is narrow on purpose: which open-source tools let one lead agent break a job into pieces and hand those pieces to specialist sub-agents, running on your own machine with your own API keys? Eight qualify. Three are finished products you install and talk to; five are frameworks you assemble the team from yourself. One of the eight is ours, and every line about it says so.
One agent that does everything is a different product category, so a single general-purpose agent is listed but never ranked first here. Sub-agents have to be something the tool actually spawns, not something you script around it.
A hosted product where the agent runs on someone else's servers does not qualify, however good it is. Self-hosting on your own server counts; so does a desktop app on your own laptop.
The licence is named in the table for every entry we could confirm from the project itself. Where we could not confirm it from the project, the table says so instead of guessing.
A tool that only works if you hold one specific vendor's subscription is a narrower answer to this question, so it is noted in the table rather than excluded outright.
This list is published by Orkas, and Orkas is one of the tools on it. Every row that describes Orkas is marked. The four rules below were fixed before the list was written, and every tool here is measured against the same four.
Finished products — install it and talk to it
A Commander agent plus named specialists — research, slides, video, office documents, UI, code — that hand one job along inside a single desktop chat. MIT licensed. It also runs Claude Code, Codex, OpenCode, OpenClaw and Hermes as local CLI backends, so an existing terminal agent becomes one member of the team.
Limits, stated plainly: there is no Orkas command line, and it cannot run a model on your own hardware — model calls go to a provider over the network, either the managed official models or your own key.
See what Orkas ships →An open-source local agent with a full CLI alongside its desktop app, dozens of MCP extensions and a wide provider list. Point it at Ollama and the model stays on your hardware — the only entry in the finished-product tier that can do that.
It is one general-purpose agent throughout, not a set of named specialists that hand work to each other. If the division of labour is the point, that is the gap.
Read the full Goose comparison →A self-hosted gateway with channel clients rather than a desktop window. Configured agents, routing, and background or nested sub-agent runs — the closest thing on this list to spawned sub-agents outside our own entry. Mature skill library, and strong at driving the machine itself.
You host it. That is an advantage if you already run a server and a cost if you do not.
Read the full OpenClaw comparison →Frameworks — you assemble the team yourself
A Python framework for multi-agent crews, MIT licensed. You write the agents, their roles and their handoffs in code. The most direct framework answer to “a lead agent coordinating specialists”, and the one with the least product around it.
Read the full CrewAI comparison →The developer framework for building LLM agents in code. Broadest ecosystem here by a distance, and the least opinionated about what a team should look like — you decide the topology.
Read the full LangChain comparison →Built around agents that keep running on their own. Pick it when the goal is genuinely unattended work and you are ready to bound the loop yourself.
Read the full AutoGPT comparison →A platform for building, evaluating and shipping AI apps to end users, self-hostable. The evaluation half is the part nothing else on this list has.
Read the full Dify comparison →An open-source visual builder you host and expose as an API. Choose it when someone who is not going to write Python still has to be able to change the flow.
Read the full Flowise comparison →| Tool | Lead agent delegates | Finished product | Your own keys |
|---|---|---|---|
| Orkas Ours | Commander plus named specialists | Yes — desktop app, MIT | Managed official models or your own provider key |
| Goose | One general agent throughout | Yes — desktop app and CLI | Wide provider list, plus Ollama on your hardware |
| OpenClaw | Routing plus background or nested sub-agent runs | Yes — but you host the gateway | Configured providers or local models |
| CrewAI | Yes, and you write the delegation | No — a Python framework, MIT | Whatever you configure in code |
| LangChain | Yes, if you build that topology | No — a developer framework | Whatever you configure in code |
| AutoGPT | Autonomous loops rather than directed handoffs | No — you run and bound the loop | Whatever you configure |
| Dify | Flows and agents you define in the platform | No — a platform you host | Configured in the platform |
| Flowise | Whatever the flow you draw does | No — a builder you host and expose | Configured in the builder |
“Finished product” is not a compliment and “framework” is not a complaint — they answer different questions. If a developer on your team is going to own this, a framework gives you more control than any product on this page. If the person who needs the work done is not that developer, a finished product is the only one of the two that gets used.
Orkas is MIT licensed and free to download for macOS and Windows. It starts with managed official models so the first run needs no provider account, and you can point it at your own key whenever you want.
Every repository fact on this page was read from the GitHub API on 11 September 2026 and read twice more the same day, so the evidence file shows which fields moved between runs and which did not. For each of the eight: the slug it actually resolves to, the licence GitHub reports, whether the repository is archived, and the latest release. The commands are in the file. Running them calls no model and costs nothing.
What this does not cover: how well any of these tools delegates, how fast it is, or how good its output is. No benchmark was run, so none should be read into the order. Two findings are recorded open rather than quietly fixed — FlowiseAI/Flowise is archived, and GitHub reports no standard licence for four of the eight.
Because leaving it out would make the list less useful and no more honest. The rules were written first, Orkas is measured against the same four as everything else, its limits are stated in its own entry, and every row about it carries a marker. Read the list knowing who wrote it.
Goose can, via Ollama. OpenClaw can use local models. Orkas cannot — its model calls go to a provider over the network, either the managed official models or your own key. If keeping the weights on your machine is the requirement, that rules Orkas out and the table says so.
Yes. Orkas runs Claude Code, Codex, OpenCode, OpenClaw and Hermes as local CLI backends, so the agent you already have becomes one member of the team rather than something you replace. Those five are the complete supported set.
Orkas is free and MIT licensed. A Commander agent and named specialists, on your machine, with managed official models or your own provider key.
