Hi 👋 Building remote-agent - an open-source free AI agent for self-hosted AI automation & workflows

[airemoteagent]

Introduce yourself

Hey GitHub 👋 I build remote-agent - an free & open-source AI agent that runs self-hosted on your own hardware (macOS, Linux, Windows, Raspberry Pi) and does real remote work and AI automation: run commands, manage files, browse the web, schedule background jobs. Every action goes through a policy engine you control, executes inside a sandbox, and lands in a tamper-evident audit log so you can see every decision and why it happened.

It's agentic AI with guardrails - bounded autonomy, prompt-injection defense, and MCP support - so an LLM can do computer-use tasks without raw shell access. Built in Node.js, MIT-licensed, one runtime dependency. Excited to share it in the open, see what breaks, and meet people working on AI security, self-hosted AI, and DevOps automation.

Links (optional)

https://remoteagent.online/
https://github.com/airemoteagent/AI-Remote-Agent

Where are you in your GitHub journey?

Maintaining a project or repo owner

And where are you going next on GitHub?

First public release of remote-agent, then a prompt-injection eval suite and (hopefully) an independent security review. Small next step: my first external contributor.

What technical skills or projects are you working on?

Node.js and AI security — policy-gated tool execution, sandboxing, prompt-injection defense, and audit trails. Also deep into self-hosted AI, agentic workflows, MCP, and DevOps automation.

Got a question for us? (optional)

For a security-focused open-source AI agent tool. How could an AI agent solve your work related challenges?

[iawaisahmd]

interesting project. for me, an ai agent would be most useful for repetitive dev work like checking builds, monitoring errors, managing deployments, and handling routine github tasks. having those actions sandboxed with clear logs would make me much more comfortable using an agent for real projects.

[airemoteagent]

The lack of transparency with these known agents is one key driver for our "Remote AI Agent". Bring back light and control to the human operator, or other LLMs. Clean log, clear boundaries and adjustable from top to bottom.

thanks for your feedback!

"The best way to run an AI agent is to build it yourself." - Einstein

[7empestx]

Since your next step is a prompt-injection eval suite and an independent review, a black-box run against the Docker image is the fastest way to get a first result: no source, no weights, just the agent answering on its HTTP route while payloads try to make it leak, override its persona, call tools it shouldn't, or exfiltrate.

I run agent-sandbox, we scan agents for prompt injection in an isolated sandbox and hand back a pass/fail report built from the agent's own responses. First cohort is free, and remote-agent is exactly the kind of self-hosted agent it was built for: https://agent.beehivewebstudio.com

[Rehanguards]

Self-hosted computer-use agents definitely need strict execution boundaries so raw shell calls aren't abused via prompt injection. Sitting right in front of the execution plane with a sub-millisecond, fail-closed policy engine prevents bad tool calls before they hit the OS layer. Great work on this—how are you currently scoping file-system and terminal tool allowlists?