Best multi-agent tools for self-hosted and local use
These tools run on your own machine or server, with your own model keys or local models, and do not require the vendor's hosted service. Licenses differ, and some are source-available rather than open source; each reason states what the record says. Data as of 2026-09-30.
Selection criteria
- Runs fully on your own hardware; any hosted service is optional.
- License terms for self-hosting are stated in the repository.
- Documents local model use or provider-neutral model configuration.
- Documents multi-agent behavior (or states clearly that it is single-agent).
- Documents how risky actions are approved or sandboxed.
Shortlist
- Langflow: MIT-licensed visual builder installable from pip, Docker or a desktop app; an Agent component can call other agents as tools, a Human Input component pauses a flow, and each flow can be served over REST, MCP or A2A.
- OpenClaw: Self-hosted Gateway (MIT text in its LICENSE file) that links chat apps to agents, with isolated multi-agent routing, sub-agents and Swarm; the README states there is no paid tier. Sandboxing is off by default.
- Hermes Agent: MIT-licensed Python agent with delegated subagents, a Kanban board for multi-process work and several terminal backends including Docker and SSH. The paid Nous Portal is optional; the default approval mode auto-approves low-risk commands.
- nanobot: MIT-licensed personal agent with one gateway for WebUI, terminal and chat apps, MCP tools, subagents and documented local models through Ollama or vLLM. Control is conversational rather than approval-based.
- Dify: Docker Compose self-hosting with a free single-workspace Community Edition; several agents can be nodes in one workflow. The license is a modified Apache 2.0 that requires a commercial license for multi-tenant operation.
- n8n: Self-hostable workflow platform with AI Agent nodes that delegate to other agents and approval gates on tools. It is fair-code (Sustainable Use License), not open source, and paid plans unlock SSO and similar features.
- AnythingLLM: MIT-licensed desktop and Docker app for private document chat with local models. Its record states it runs one tool-using agent per chat and is not a multi-agent orchestrator; listed for local-first use.
Shortlist facts
| Tool | Type | Languages | License | Pricing | Pattern | MCP | A2A | AG-UI | Stars |
|---|---|---|---|---|---|---|---|---|---|
| Langflow | Platform | Python, TypeScript | MIT | Open source, free | Supervisor | Yes evidence | Yes evidence | Yes evidence | 155,389 |
| OpenClaw | Harness | TypeScript | unknown | Open source, free | Supervisor | Yes evidence | Yes evidence | Unknown | 390,840 |
| Hermes Agent | Harness | Python | MIT | Open core | Supervisor | Yes evidence | Yes evidence | Unknown | 250,234 |
| nanobot | Harness | Python, TypeScript | MIT | Open source, free | Supervisor | Yes evidence | Unknown | Unknown | 48,696 |
| Dify | Platform | Python, TypeScript, JavaScript | unknown | Open core | Graph | Yes evidence | Unknown | Unknown | 157,573 |
| n8n | Platform | TypeScript, JavaScript, Python | unknown | Open core | Supervisor | Yes evidence | Unknown | Unknown | 206,347 |
| AnythingLLM | Platform | JavaScript | MIT | Open core | Other | Yes evidence | Unknown | Unknown | 66,618 |
FAQ
Which of these are open source in the strict sense?
Langflow, Hermes Agent, nanobot and AnythingLLM are MIT (OpenClaw's LICENSE file also contains MIT text). Dify and n8n use their own licenses with restrictions, so GitHub reports no standard license for them.
Can they run without any cloud model?
Each record lists local model support where the docs state it, for example nanobot and AnythingLLM with Ollama. Quality depends on the local model, which these pages do not test.
Are self-hosted agents safe by default?
Not by default. OpenClaw runs main-session tools on the host until sandboxing is configured, and Hermes' default mode lets a model auto-approve low-risk commands. Read the HITL section of each tool page before leaving an agent running.