Hivelock: Self-Hosted AI Agent Platform with Human Approvals

Hi Product Hunt 👋

We built Hivelock because every team we spoke to hit the same wall with AI agents. The demo works, then someone asks: what stops it from doing something we didn’t intend, and who checks its work?

Agent projects tend to fail in four ways, so we built a control for each one:

Acting beyond intent: each agent gets only the actions you choose, and credentials never reach the model. Reading runs freely; anything that changes data waits for a person.
Runaway cost: per-run step and token limits, a loop guard that stops repeated calls, and a monthly AI allowance with a warning at 80%.
Silent breakage: save tests for an agent. When someone changes it, the tests run and live workflows keep off the new version until they pass.
Missing oversight: pause everything with one button or turn off a single agent, with a tamper-evident audit log of every change and approval.

Under that sits a visual workflow builder, 74+ templates, 26+ connectors (including remote MCP servers), and support for any OpenAI-compatible model, including local ones.

It’s self-hosted and the core is open source (AGPL-3.0). The free Community edition has the same security as the paid one: encryption with a key for each workspace, row-level isolation and two-step sign-in.

We’re in early access and honest about it. We test it ourselves with Strix, but it hasn’t had a third-party pentest yet.

Feedback that would help most:

If you’ve tried agents and pulled them back, what made you stop?
Is approving each action the right default, or would you want it looser or stricter?
Which connector or template should we add next?

🔗 Code: https://github.com/shichifukutekx/hivelock
🌐 Site: https://hivelock.shichifukutekx.ai/

Thanks for taking a look. We’ll be here all day answering questions.