Comparison
An honest comparison. We build Pinchy, one of the tools on this list, so judge accordingly. We have tried to be scrupulously fair, we link every competitor, and we tell you when one of them is the better choice.
"Self-hosted AI agent platform" has become an umbrella term for at least four different product categories, and most comparison lists mix them up. So this one starts by separating them, because the right pick depends entirely on which job you are hiring for. Corrections welcome via GitHub issue.
The hard part of running AI agents is not building them, it is controlling what they do. Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, citing unclear value, rising costs, and weak governance. Databricks' 2026 enterprise research ranks governance as the number-one buyer priority for AI agents, ahead of model choice and deployment target. And under the EU AI Act, high-risk deployments must keep logs, preserve human oversight, and stay auditable from December 2027, with fines reaching 7% of global turnover. That is why the comparison below weighs free-tier permissions and audit trails, not just feature counts.
Windmill, best for developer-built workflow automation. A mature script and workflow engine (AGPLv3 core, around 16.7k GitHub stars, a fast release cadence) with a genuinely strong permission model in the free tier: workspace roles, path-based ACLs, folders. AI agents exist as flow steps with tool calling, MCP support, and a chat mode. Know the community edition limits before you commit: 50 users, 4 groups, SSO capped at 10 users, and durable, exportable audit logs require the Enterprise Edition (the docs state 14-day audit retention for CE). If your team thinks in scripts and scheduled flows, pick Windmill.
Onyx, best for permission-aware company knowledge search. Formerly Danswer (MIT core with enterprise-licensed directories, around 30k stars, $10M seed funding). Chat over 50+ connectors with strong RAG, agents with actions and MCP, and explicit local-LLM support (Ollama, vLLM, LiteLLM). Its standout feature, syncing document permissions from source systems like Google Drive or Confluence, is enterprise-licensed for self-hosted deployments, as are SSO, RBAC, and user groups. If your job is "let the team ask questions across our documents, respecting who may see what", pick Onyx.
Pinchy, best for governed agents that act, not just answer. Our product, so judge accordingly: a self-hosted platform (AGPL-3.0) where end users chat with agents through a web UI or Telegram, and admins control everything the agents may do. New agents start with zero tools; each tool is enabled explicitly per agent. Every agent action lands in an audit trail whose entries are individually HMAC-signed, so tampering is detectable. To our knowledge, neither Windmill nor Onyx document cryptographic signing of their logs. The free tier has no user limit and includes the permission core (allow-list tools per agent) and the signed audit trail; groups, agent visibility, and per-user analytics are part of the paid Pro tier (€99/month for up to 10 users, prices public). The honest caveats: Pinchy is young, the integration list is short (Odoo ERP, Gmail, Telegram, web search, documents), there is no compliance certification yet, and granular RBAC is on the roadmap. If your job is "let the team use AI agents on real business systems without losing control", that is exactly what we built.
n8n, best for visual no-code automation. The Zapier-class tool of self-hosting, with AI agent nodes. Important nuance: n8n is fair-code under the Sustainable Use License, not OSI open source. Internal use is fine, embedding or reselling is not.
LibreChat (MIT) is the cleanest multi-provider chat UI with agent and RAG extensions; a chat frontend rather than a governance platform. AnythingLLM (MIT) is the easiest private chat-with-your-documents setup for individuals and small teams. Dify and Flowise are visual builders for LLM apps (check their license carve-outs). CrewAI (MIT) is a Python framework, not a product; you build and operate the platform yourself.
| Category | License (core) | Free-tier governance | Local LLMs | Best for | |
|---|---|---|---|---|---|
| Windmill | workflow engine | AGPLv3* | roles + item ACLs (50 users, 4 groups; durable audit = EE) | via custom endpoints | developer automation |
| Onyx | knowledge search | MIT* | basic auth/roles (RBAC, SSO, permission-sync = EE) | yes, explicit | company-wide Q&A |
| Pinchy | governed team agents | AGPL-3.0 | allow-list tools per agent + signed audit, no user limit (groups/visibility = Pro) | yes (Ollama) | agents acting in business systems |
| n8n | automation | fair-code (SUL) | RBAC/SSO paid | yes | visual workflows |
| LibreChat | chat UI | MIT | basic | yes | multi-provider chat |
* With enterprise-licensed components; see each project's license file. All figures as of June 2026; prices and limits change, so check the vendors' pages.
Ask what the agent is supposed to touch. Documents only: Onyx. Scripts and scheduled jobs, built by developers: Windmill. Business systems, operated by non-developers, with someone accountable for what the agents did: Pinchy. If you would rather assemble your own, the builders and frameworks above are the place to start.
Pinchy is free and open source to self-host. Groups, agent visibility, and analytics come with a flat Pro tier.
Or email us: info@heypinchy.com